diff --git a/AGENTS.md b/AGENTS.md
index 5c4d661..d16a742 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -1,5 +1,16 @@
 # PyAutoPulse — agent instructions
 
+## Profiling campaign front door
+
+Read `CHECKIN.md` for all-campaign updates and one-chat operation. Pulse now owns
+profiling campaign intent and pending tasks in `campaigns.yaml` and `tasks/`,
+including prompts migrated from Mind (`migration.yaml`). The board displays
+the editable check-in prompt, active campaigns, open tasks, then measurements.
+Mind retains bounded development issue/PR lifecycle and repository claims;
+project repos retain measurement evidence. This explicit ownership update takes
+precedence over older descriptions of Pulse as only a summary reader.
+
+
 This file is for AI coding agents (Claude Code, Codex, Cursor, etc.) and humans
 discovering this repository. PyAutoPulse is the **Pulse** organ of the PyAuto
 organism — where the organism feels how fast it runs. It is the cross-project
diff --git a/CHECKIN.md b/CHECKIN.md
new file mode 100644
index 0000000..b5cbac9
--- /dev/null
+++ b/CHECKIN.md
@@ -0,0 +1,58 @@
+# All profiling work in one chat
+
+The editable prompt at the top of the dashboard is the entry point. Direction
+supplied before or after it has the same meaning: focus that campaign or assess
+that idea while retaining the overall sweep. No direction means check all open
+campaigns and tasks. This static page does not run jobs or monitor them itself.
+
+## Check-in procedure
+
+1. Fetch current Pulse, Mind and relevant project branches; read their agent
+   instructions. Read `campaigns.yaml`, task files and `migration.yaml` when an
+   old Mind path is referenced. Pulse is the canonical profiling queue.
+2. Read each campaign's current project wiki/ledger, results and issue/PR state.
+   Evidence URLs pin the last reviewed source; inspect the current default
+   branch for subsequent changes too. Consult Brain's profiling conductor for
+   scientific interpretation. With HPC access, use project pull/status commands;
+   otherwise report which jobs/results remain unverified. Never infer completion
+   from an old job id or absent data.
+3. Compare with the last check-in and report one table: changes, current state,
+   blockers, next bounded step. Keep timing, correctness, compile cost, memory
+   and hardware distinct. Respect task contracts, human decisions, release gates
+   and Mind's current repository claims.
+4. Incorporate user direction into priorities and suggested next steps. An idea
+   is a hypothesis until tested. Choose one bounded phase before execution and
+   use existing development/compute authorization rules. The check-in itself
+   authorizes no job submission, default/baseline change, release or merge.
+5. Update `campaigns.yaml` with verified facts, reviewed dates, next steps and
+   source links. Preserve parked/blocked work. Append decisions and their
+   evidence to the task or project ledger. Record `last_checkin` as an ISO UTC
+   timestamp after the sweep, noting unavailable sources in the affected tasks.
+   A board refresh alone must never stamp a check-in.
+6. Run `bin/pyauto-pulse board --offline` (or live ingest when available), then
+   `bin/pyauto-pulse check --offline`; persist through the repo development
+   workflow. Report the PR and whether it merged. Continue subsequent profiling
+   requests in this conversation; a fresh chat resumes from these records.
+
+## Ownership and development handoff
+
+Pulse owns campaign intent and pending tasks. File new profiling work under
+`tasks/` with a row in `campaigns.yaml`. Each task has a stable ID, campaign,
+title, status, priority, path, reviewed date and next step. Campaigns carry a
+title, status, reviewed date, next step and evidence URL. Completed/superseded
+rows remain in the ledger but leave the open table. Parked and blocked work
+stays visible. A campaign is not a job: active does not mean running.
+
+For a library/workspace implementation phase, file a small Mind development
+prompt referencing the Pulse task, then use start_dev and the normal issue/PR
+and claim lifecycle. Reuse existing issues (including autolens_profiling#362 for
+timing-noise audit). Do not copy entire campaigns into Mind or bulk issue their
+phases. After close-out, link the completion record from the Pulse task and
+update its next step. Mind retains repository claims and implementation PR
+lifecycle; mixed science/library work stays there with Pulse links.
+
+Projects retain detailed campaign evidence, producers, results, pins and drift
+policy. Pulse scheduling metadata never changes measurement qualification or
+Heart readiness. Migration preserves original text and provenance; historical
+headers inside those texts are not a live queue. Resolve old Mind paths via
+`migration.yaml`; use the pinned source commit for unmigrated historical paths.
diff --git a/README.md b/README.md
index 504683c..232199f 100644
--- a/README.md
+++ b/README.md
@@ -1,5 +1,11 @@
 # PyAutoPulse
 
+The [campaign control room](CHECKIN.md) provides one editable check-in prompt,
+active campaigns and Pulse-owned tasks above the detailed profiling evidence.
+`campaigns.yaml` is the scheduling ledger; `migration.yaml` records original Mind
+prompts and their source commit. Board refreshes never change task status or stamp
+a check-in. Run `bin/pyauto-pulse board --offline` after ledger edits.
+
 The **Pulse** organ of the [PyAuto organism](https://github.com/PyAutoLabs/PyAutoBrain/blob/main/ORGANISM.md):
 where the organism feels how fast it runs.
 
diff --git a/REFERENCE.md b/REFERENCE.md
index e83aff7..86ea9a7 100644
--- a/REFERENCE.md
+++ b/REFERENCE.md
@@ -1,5 +1,11 @@
 # PyAutoPulse — reference
 
+The [campaign control room](CHECKIN.md) provides one editable check-in prompt,
+active campaigns and Pulse-owned tasks above the detailed profiling evidence.
+`campaigns.yaml` is the scheduling ledger; `migration.yaml` records original Mind
+prompts and their source commit. Board refreshes never change task status or stamp
+a check-in. Run `bin/pyauto-pulse board --offline` after ledger edits.
+
 The registry schema, the `profiling-summary` v1 contract as this organ reads
 it, and the formats of the receipts, snapshots, board markers and cockpit feed.
 Prose boundaries live in [AGENTS.md](AGENTS.md); the design is
diff --git a/campaigns.yaml b/campaigns.yaml
new file mode 100644
index 0000000..a406f30
--- /dev/null
+++ b/campaigns.yaml
@@ -0,0 +1,285 @@
+version: 1
+last_checkin: null
+campaigns:
+- id: point-source-image-cpu
+  title: Point-source image plane · CPU
+  status: active
+  updated: '2026-10-03'
+  next: Phases 1–4 shipped; reconcile carried cleanup and test-placement work.
+  evidence: https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/point_source_image_plane_cpu.md
+- id: point-source-image-gpu
+  title: Point-source image plane · A100
+  status: needs-decision
+  updated: '2026-10-03'
+  next: Measure an image-plane fit for the admission decision; retain forward-gradient correctness findings.
+  evidence: https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/point_source_gpu_breakdown.md
+- id: point-source-source
+  title: Point-source source plane
+  status: parked
+  updated: '2026-10-03'
+  next: Core complete; gradient-sampler admission required before blackjax work. Warm-up measurement fix remains
+    open.
+  evidence: https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/point_source_source_plane.md
+- id: cluster-pointsolver
+  title: Cluster PointSolver robustness and performance
+  status: active
+  updated: '2026-10-03'
+  next: Reconcile the expanded robustness contract; select one bounded phase and current representative baseline.
+  evidence: https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/cluster_pointsolver.md
+- id: interferometer
+  title: Interferometer likelihood and streaming
+  status: active
+  updated: '2026-10-03'
+  next: Collect the final CPU matrix cell; scope streaming scaling, curvature reuse and FFT levers separately.
+  evidence: https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/interferometer_likelihood.md
+- id: linear-solver
+  title: Linear-solver accuracy and cost
+  status: blocked
+  updated: '2026-10-03'
+  next: Verify release containing Array#595 and RAL sync before GPU/vmap parity and timing.
+  evidence: https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/linear_solver_accuracy.md
+- id: imaging
+  title: Production imaging and likelihood breakdown
+  status: active
+  updated: '2026-10-03'
+  next: Reconcile measured GPU/CPU coverage, then profile production over-sampling and remaining costs.
+  evidence: https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/post_certified_breakdown.md
+- id: measurement-tools
+  title: Measurement reliability and profiling tools
+  status: active
+  updated: '2026-10-03'
+  next: Audit timing noise; repair warm-up, cumulative accounting, memo blindness and compile-cell construction.
+  evidence: https://github.com/PyAutoLabs/PyAutoMind/tree/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft
+- id: autofit
+  title: PyAutoFit profiling bootstrap
+  status: needs-decision
+  updated: '2026-10-03'
+  next: Confirm new repository creation before porting baseline packages; no empty producer registration.
+  evidence: https://github.com/PyAutoLabs/PyAutoMind/tree/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft
+- id: critical-curves
+  title: Critical curves and evaluation grids
+  status: blocked
+  updated: '2026-10-03'
+  next: Galaxy#646 and profiling#365 merged; workspace#343 remains release-gated in Mind.
+  evidence: https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/critical_curves.md
+- id: certified-solver
+  title: Certified positive solver
+  status: parked
+  updated: '2026-10-03'
+  next: Retain the parked opt-in solver study; read the project ledger before proposing new work.
+  evidence: https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/certified_positive_solver.md
+tasks:
+- id: runtime_cell_single_jit_gpu_warmup
+  campaign: point-source-source
+  title: Runtime cells' A100 `single_jit` includes the post-compile warm-up — source-plane 0.642 ms vs a steady
+    0.267 ms on the same node
+  status: needs-decision
+  priority: low
+  path: tasks/runtime_cell_single_jit_gpu_warmup.md
+  updated: '2026-10-03'
+  next: Read the task and resolve its explicit human decision before execution.
+- id: timing_noise_audit
+  campaign: measurement-tools
+  title: Audit timing tests and profiling gates for measurement noise
+  status: ready
+  priority: high
+  path: tasks/timing_noise_audit.md
+  updated: '2026-10-03'
+  next: Read the task and select the next bounded step; check current repository claims.
+  issue: https://github.com/PyAutoLabs/autolens_profiling/issues/362
+- id: mge_likelihood_breakdown_steps_are_cumulative_an
+  campaign: measurement-tools
+  title: MGE likelihood_breakdown steps are cumulative and `linear_gaussians` is reported as 0
+  status: ready
+  priority: low
+  path: tasks/mge_likelihood_breakdown_steps_are_cumulative_an.md
+  updated: '2026-10-03'
+  next: Read the task and select the next bounded step; check current repository claims.
+- id: gradient_cost_probe
+  campaign: measurement-tools
+  title: 'A gradient-cost probe: forward vs `value_and_grad` ms/eval and a strict FD check, on any registry cell'
+  status: ready
+  priority: low
+  path: tasks/gradient_cost_probe.md
+  updated: '2026-10-03'
+  next: Read the task and select the next bounded step; check current repository claims.
+- id: numba_breakdown_harness_memo_blind
+  campaign: measurement-tools
+  title: 'Numba breakdown harness: perturb the instance so the operated-matrix memo cannot hide a step'
+  status: ready
+  priority: medium
+  path: tasks/numba_breakdown_harness_memo_blind.md
+  updated: '2026-10-03'
+  next: Read the task and select the next bounded step; check current repository claims.
+- id: search_settings_estimation_infrastructure
+  campaign: measurement-tools
+  title: Search settings-estimation + profiling infrastructure (n_starts / batch_size / n_batch)
+  status: ready
+  priority: normal
+  path: tasks/search_settings_estimation_infrastructure.md
+  updated: '2026-10-03'
+  next: Read the task and select the next bounded step; check current repository claims.
+- id: jax_compile_probe_needs_own_cell_builder
+  campaign: measurement-tools
+  title: jax_compile/probe.py lost its cell builder with the searches tier — give profiling its own
+  status: ready
+  priority: normal
+  path: tasks/jax_compile_probe_needs_own_cell_builder.md
+  updated: '2026-10-03'
+  next: Read the task and select the next bounded step; check current repository claims.
+- id: mass_field_flat_adoption_science_repos
+  campaign: measurement-tools
+  title: 'Remaining code consumers move to the flat `fields=` form: profiling, inference, JOSS benchmarks, Reduce
+    prototypes'
+  status: superseded
+  priority: normal
+  path: tasks/mass_field_flat_adoption_science_repos.md
+  updated: '2026-10-03'
+  next: Historical umbrella; shipped work retained. Do not issue it again.
+- id: mge_nnls_fix_pyautoarray_571_slam_60
+  campaign: linear-solver
+  title: 'Linear-solver programme phase 3: GPU/vmap/A100 timing and parity rows for the solver corpus'
+  status: blocked
+  priority: medium
+  path: tasks/mge_nnls_fix_pyautoarray_571_slam_60.md
+  updated: '2026-10-03'
+  next: a release shipping the phase-2 fix (PyAutoArray#595, merged 2026-09-30 pending release; record complete/2026/09/raw-pdip-forward-polish.md)
+    AND the RAL mirror synced via HPCPullPyAuto
+- id: autofit_profiling_bootstrap
+  campaign: autofit
+  title: 'autofit_profiling: bootstrap the repo + general PyAutoFit profiling epic'
+  status: needs-decision
+  priority: normal
+  path: tasks/autofit_profiling_bootstrap.md
+  updated: '2026-10-03'
+  next: Read the task and resolve its explicit human decision before execution.
+- id: cluster_pointsolver_speed
+  campaign: cluster-pointsolver
+  title: Cluster PointSolver — robustness, analysis settings and performance
+  status: ready
+  priority: normal
+  path: tasks/cluster_pointsolver_speed.md
+  updated: '2026-10-03'
+  next: Read the task and select the next bounded step; check current repository claims.
+- id: imaging_production_over_sampling
+  campaign: imaging
+  title: Profile imaging pixelizations at production over-sampling
+  status: ready
+  priority: normal
+  path: tasks/imaging_production_over_sampling.md
+  updated: '2026-10-03'
+  next: Read the task and select the next bounded step; check current repository claims.
+- id: interferometer_decision_matrix_last_cell
+  campaign: interferometer
+  title: Interferometer decision matrix — fill the last cell (CPU rect 39² at alma_high r5.0) and flip the wiki
+    row to shipped
+  status: needs-decision
+  priority: normal
+  path: tasks/interferometer_decision_matrix_last_cell.md
+  updated: '2026-10-03'
+  next: Read the task and resolve its explicit human decision before execution.
+- id: interferometer_fixed_mapper_curvature_preload
+  campaign: interferometer
+  title: 'Interferometer fixed-mapper searches: reuse the W~ curvature matrix across likelihood calls via `preloads.curvature_matrix`'
+  status: needs-slicing
+  priority: normal
+  path: tasks/interferometer_fixed_mapper_curvature_preload.md
+  updated: '2026-10-03'
+  next: Scope one bounded phase from the full task contract.
+- id: interferometer_nnls_memo_scattered_stream_guard
+  campaign: interferometer
+  title: 'fnnls warm-start memo: stop it slowing scattered evaluation streams (interferometer CPU, autolens_profiling#332)'
+  status: ready
+  priority: medium
+  path: tasks/interferometer_nnls_memo_scattered_stream_guard.md
+  updated: '2026-10-03'
+  next: Read the task and select the next bounded step; check current repository claims.
+- id: interferometer_streaming_scaling
+  campaign: interferometer
+  title: 'Campaign: interferometer streaming (array-free) vs in-memory — memory and time scaling to 2e8 visibilities'
+  status: ready
+  priority: high
+  path: tasks/interferometer_streaming_scaling.md
+  updated: '2026-10-03'
+  next: Read the task and select the next bounded step; check current repository claims.
+- id: interferometer_w_tilde_fft_size_levers
+  campaign: interferometer
+  title: 'Interferometer W~ curvature matrix is FFT-bound on the mask extent: pruned padded FFT and real-space pixel
+    scale on the A100'
+  status: ready
+  priority: normal
+  path: tasks/interferometer_w_tilde_fft_size_levers.md
+  updated: '2026-10-03'
+  next: Read the task and select the next bounded step; check current repository claims.
+- id: point_solver_profiling_cells
+  campaign: cluster-pointsolver
+  title: 'PointSolver profiling cells: lensed quasar → cluster runtime tier → single/multi-source → multiplane'
+  status: ready
+  priority: normal
+  path: tasks/point_solver_profiling_cells.md
+  updated: '2026-10-03'
+  next: Read the task and select the next bounded step; check current repository claims.
+- id: point_source_image_plane_gpu_breakdown
+  campaign: point-source-image-gpu
+  title: 'Point-source A100 speed-up campaign: profile and optimize with the shared breakdown'
+  status: needs-decision
+  priority: high
+  path: tasks/point_source_image_plane_gpu_breakdown.md
+  updated: '2026-10-03'
+  next: Read the task and resolve its explicit human decision before execution.
+- id: point_source_source_plane_chi_squared_speed
+  campaign: point-source-source
+  title: Point-source source-plane chi-squared speed-up campaign — remaining candidates (phases 1–2e shipped)
+  status: blocked
+  priority: low
+  path: tasks/point_source_source_plane_chi_squared_speed.md
+  updated: '2026-10-03'
+  next: Await gradient-sampler admission evidence; core campaign already complete.
+- id: pointsolver_cpu_speed_campaign_remainder
+  campaign: point-source-image-cpu
+  title: Point-source (single-source) CPU campaign — carried leftovers and completion evidence
+  status: ready
+  priority: low
+  path: tasks/pointsolver_cpu_speed_campaign_remainder.md
+  updated: '2026-10-03'
+  next: Read the task and select the next bounded step; check current repository claims.
+- id: post_certified_solver_likelihood_breakdown
+  campaign: imaging
+  title: Assess remaining likelihood bottlenecks after certified solver integration on CPU and GPUs
+  status: needs-slicing
+  priority: normal
+  path: tasks/post_certified_solver_likelihood_breakdown.md
+  updated: '2026-10-03'
+  next: Scope one bounded phase from the full task contract.
+- id: profile_lens_aggregator_needs_config_dir
+  campaign: measurement-tools
+  title: 'profile_lens_aggregator.py cannot run from the autolens_workspace_developer root: no config/ directory'
+  status: ready
+  priority: low
+  path: tasks/profile_lens_aggregator_needs_config_dir.md
+  updated: '2026-10-03'
+  next: Read the task, reconcile current evidence and select one bounded step.
+- id: gradient_pixelization_pin_residual_drift
+  campaign: measurement-tools
+  title: 'jax_profiling/gradient/imaging/pixelization.py: 3.2% of its pin move is unattributed'
+  status: ready
+  priority: normal
+  path: tasks/gradient_pixelization_pin_residual_drift.md
+  updated: '2026-10-03'
+  next: Read the task, reconcile current evidence and select one bounded step.
+- id: pair_jax_xla_env_vars_with_measured
+  campaign: measurement-tools
+  title: Pair JAX/XLA env vars with measured compile and run times, per backend
+  status: ready
+  priority: high
+  path: tasks/pair_jax_xla_env_vars_with_measured.md
+  updated: '2026-10-03'
+  next: Read the task, reconcile current evidence and select one bounded step.
+- id: quick_update_plotting_cost
+  campaign: imaging
+  title: Quick-update plotting cost — minutes per update, and it is not JAX compile
+  status: ready
+  priority: medium
+  path: tasks/quick_update_plotting_cost.md
+  updated: '2026-10-03'
+  next: Read the task, reconcile current evidence and select one bounded step.
diff --git a/dashboard.html b/dashboard.html
index 0e7823d..c5707a0 100644
--- a/dashboard.html
+++ b/dashboard.html
@@ -34,8 +34,21 @@
 background:var(--bg);color:var(--fg);border:1px solid var(--line);border-radius:6px;
 padding:3px 6px;cursor:pointer;word-break:break-all}
 button.copied{border-color:var(--ok)}
+.checkin textarea{display:block;width:100%;font:inherit;font-size:14px;line-height:1.5;background:var(--card);color:var(--fg);border:1px solid var(--line);border-radius:8px;padding:12px;margin:8px 0;resize:vertical}
+#copy-checkin{background:var(--accent);color:var(--bg);border:0;border-radius:6px;padding:10px 16px;cursor:pointer;font-weight:600}
+#copy-status{margin-left:12px}
 code{color:var(--accent);word-break:break-all}
-</style></head><body><header class='hero'><h1>PyAuto<span>Pulse</span></h1><p>Measure. Trace. Compare.</p></header><main><p class='lede'>The cross-project view of the organism&#x27;s profiling. Each <code>&lt;lib&gt;_profiling</code> project repo runs its producers, keeps its results and drift policy, and publishes a <code>profiling-summary</code> file. This page reads each file at one resolved commit, validates the exchange contract and shows what the producer says: it recomputes no ratio, combines no timings whose axis or hardware differs, ranks nothing across projects and issues no verdict. Judgment belongs to the Brain&#x27;s profiling conductor (<code>/profiling triage</code>).</p><div class='stats'><div><b>1</b><span>Projects</span></div><div><b>159</b><span>Records</span></div><div><b>145</b><span>Comparisons</span></div><div><b>0</b><span>Drifted</span></div><div><b>0</b><span>Refused pairs</span></div><div><b>0</b><span>Cached</span></div><div><b>0</b><span>Failed</span></div></div><h2 id='projects'>Projects</h2><div class='tablewrap'><table><thead><tr><th>Project</th><th>Scope</th><th>Evidence</th><th>Last fetch</th><th>Coverage</th><th>Integrity</th><th>Freshness</th><th>Qualification</th><th>Links</th></tr></thead><tbody><tr><td><a href='#lens'>lens</a> <span class='muted'>autolens_profiling</span></td><td>release-runtime</td><td>2026-08-17</td><td>2026-10-03 @ efcece3e</td><td>145 series · 159 points · 4 releases · 61 cells · 6 excluded</td><td class='ok'>ok</td><td>freshness policy unspecified · age 47 d at fetch 2026-10-03</td><td>2/159 records · 0/145 comparisons qualified</td><td><a href='https://pyautolabs.github.io/autolens_profiling/'>dashboard</a> · <a href='https://github.com/PyAutoLabs/autolens_profiling'>repo</a> · <a href='https://github.com/PyAutoLabs/PyAutoPulse/blob/main/receipts/lens.json'>receipt</a></td></tr></tbody></table></div><p class='muted'>Three notions are kept apart: <b>integrity</b> is whether the file was fetched and meets the contract; <b>freshness</b> is the evidence&#x27;s age against the producer&#x27;s own <code>valid_until</code> (none declared → &quot;freshness policy unspecified&quot;); <b>qualification</b> is the producer&#x27;s own flag on each record and comparison. Healthy transport never makes unknown quality green.</p><h2 id='lens'>lens · autolens_profiling</h2><p class='lede'>Project <code>autolens_profiling</code>, scope <code>release-runtime</code>, read from <a href='https://github.com/PyAutoLabs/autolens_profiling'>PyAutoLabs/autolens_profiling</a> <code>dashboard/summary.json</code> at <a href='https://github.com/PyAutoLabs/autolens_profiling/tree/efcece3ea39c622e62cc61518f58394ee3986d4c'><code>efcece3e</code></a>; producer revision <code>227b5c91</code>; generated 2026-09-30T19:43:00Z; policy <code>runtime-drift-2x-1ms</code>. <a href='https://pyautolabs.github.io/autolens_profiling/'>Project dashboard</a> · <a href='https://github.com/PyAutoLabs/PyAutoPulse/blob/main/receipts/lens.json'>receipt</a>.</p><div class='notions'><div><b>Integrity</b><span class='ok'>ok</span></div><div><b>Freshness</b>freshness policy unspecified · age 47 d at fetch 2026-10-03 (evidence 2026-08-17)<span class='muted' data-age-from='2026-08-17T00:00:00Z'></span></div><div><b>Qualification</b>2/159 records · 0/145 comparisons qualified</div></div><h3>Limitations (the producer's own)</h3><ul><li>runtime axis only: compile times, VRAM and per-step component timings are not in this feed</li><li>evidence_updated_at is the newest release date encoded in a measured library version; result rows do not record measurement wall-clock time</li><li>valid_until is null: this project declares no freshness policy for release trends</li><li>vmap_per_call_s is null where a row carries no vmap block</li><li>rows without a device.provenance block (pre-autolens_profiling#342) are unqualified</li><li>6 row(s) above the load-average cap 8 are listed under coverage.excluded and never enter records</li></ul><h3>Excluded by the producer</h3><ul><li><code>breakdown:interferometer/sdp81/delaunay/sdp81:hpc_a100_fp64@2026.8.17.1</code>: loadavg 65.7 above the cap 8 (<a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/interferometer/sdp81/delaunay_hpc_a100_fp64.json'>evidence</a>)</li><li><code>breakdown:interferometer/sdp81/delaunay_numba/sdp81:hpc_ral_cpu_fp64@2026.8.17.1</code>: loadavg 68.3 above the cap 8 (<a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/interferometer/sdp81/delaunay_numba_hpc_ral_cpu_fp64.json'>evidence</a>)</li><li><code>breakdown:interferometer/sdp81/mge/sdp81:hpc_a100_fp64@2026.8.17.1</code>: loadavg 55.7 above the cap 8 (<a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/interferometer/sdp81/mge_hpc_a100_fp64.json'>evidence</a>)</li><li><code>breakdown:interferometer/sdp81/mge/sdp81:hpc_a100_fp64_sparse@2026.8.17.1</code>: loadavg 55.7 above the cap 8 (<a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/interferometer/sdp81/mge_hpc_a100_fp64_sparse.json'>evidence</a>)</li><li><code>breakdown:interferometer/sdp81/pixelization/sdp81:hpc_a100_fp64@2026.8.17.1</code>: loadavg 65.7 above the cap 8 (<a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/interferometer/sdp81/pixelization_hpc_a100_fp64.json'>evidence</a>)</li><li><code>breakdown:interferometer/sdp81/pixelization_numba/sdp81:hpc_ral_cpu_fp64@2026.8.17.1</code>: loadavg 68.3 above the cap 8 (<a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/interferometer/sdp81/pixelization_numba_hpc_ral_cpu_fp64.json'>evidence</a>)</li></ul><h3>Comparisons on cpu · cpu · float64</h3><div class='tablewrap'><table><thead><tr><th>Comparison</th><th>Policy</th><th>Axis · metric</th><th>Baseline → candidate</th><th>Ratio (producer)</th><th>Status</th><th>Qualified</th><th>Reasons</th><th>Triage</th></tr></thead><tbody><tr><td><code>runtime:imaging/delaunay/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/runtime/imaging/delaunay/hst/comparison.json'>2026.7.6.649 [local_cpu_fp64]</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/runtime/imaging/delaunay/delaunay_likelihood_summary_hst_v2026.8.17.1.json'>2026.8.17.1</a></td><td class='num'>0.27</td><td>improved</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/delaunay/hst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/hst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/hst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/delaunay/jwst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/jwst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/jwst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/delaunay/jwst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/jwst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/jwst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/delaunay_nn/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay_nn/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay_nn/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/delaunay_numba/euclid:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay_numba/euclid:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay_numba/euclid:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/delaunay_numba/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay_numba/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay_numba/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/mge/ao:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/ao:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/ao:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/mge/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/runtime/imaging/mge/hst/comparison.json'>2026.7.6.649 [local_cpu_fp64]</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/runtime/imaging/mge/mge_likelihood_summary_hst_v2026.8.17.1.json'>2026.8.17.1</a></td><td class='num'>0.671</td><td>flat</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/mge/hst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/hst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/hst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/mge/jwst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/jwst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/jwst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/mge/jwst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/jwst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/jwst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/mge_mass_jax/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge_mass_jax/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge_mass_jax/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/pixelization/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/runtime/imaging/pixelization/hst/comparison.json'>2026.7.6.649 [local_cpu_fp64]</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/runtime/imaging/pixelization/pixelization_likelihood_summary_hst_v2026.8.17.1.json'>2026.8.17.1</a></td><td class='num'>0.232</td><td>improved</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/pixelization/hst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/hst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/hst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/pixelization/jwst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/jwst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/jwst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/pixelization/jwst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/jwst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/jwst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/pixelization_numba/euclid:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization_numba/euclid:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization_numba/euclid:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/pixelization_numba/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization_numba/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization_numba/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/pixelization_numba_mge_mass/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization_numba_mge_mass/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization_numba_mge_mass/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:interferometer/delaunay/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:interferometer/delaunay/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:interferometer/delaunay/sma:local_cpu_fp64</button></td></tr><tr><td><code>runtime:interferometer/mge/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:interferometer/mge/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:interferometer/mge/sma:local_cpu_fp64</button></td></tr><tr><td><code>runtime:interferometer/pixelization/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:interferometer/pixelization/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:interferometer/pixelization/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:cluster/image_plane:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/cluster/image_plane_breakdown_v2026.7.23.1.json'>2026.7.23.1</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/cluster/image_plane_local_cpu_fp64.json'>2026.8.17.1</a></td><td class='num'>0.469</td><td>improved</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage breakdown:cluster/image_plane:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:cluster/image_plane:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:cluster/source_plane:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/cluster/source_plane_breakdown_v2026.7.6.649.json'>2026.7.6.649</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/cluster/source_plane_breakdown_v2026.7.23.1.json'>2026.7.23.1</a></td><td class='num'>1.469</td><td>flat</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage breakdown:cluster/source_plane:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:cluster/source_plane:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/delaunay/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/delaunay_breakdown_hst_v2026.7.6.649.json'>2026.7.6.649</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/delaunay_breakdown_hst_v2026.8.17.1.json'>2026.8.17.1</a></td><td class='num'>0.331</td><td>improved</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay/hst:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/delaunay/hst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/delaunay_breakdown_hst_v2026.5.29.4_sparse.json'>2026.5.29.4</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/delaunay_breakdown_hst_v2026.7.6.649_sparse.json'>2026.7.6.649</a></td><td class='num'>1.108</td><td>flat</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay/hst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay/hst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>breakdown:imaging/delaunay_nn/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay_nn/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay_nn/hst:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/delaunay_numba/euclid:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay_numba/euclid:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay_numba/euclid:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/delaunay_numba/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay_numba/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay_numba/hst:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/mge/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/mge_breakdown_hst_v2026.7.6.649.json'>2026.7.6.649</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/mge_breakdown_hst_v2026.8.17.1.json'>2026.8.17.1</a></td><td class='num'>0.225</td><td>improved</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/mge/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/mge/hst:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/pixelization/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/pixelization_breakdown_hst_v2026.7.6.649.json'>2026.7.6.649</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/pixelization_breakdown_hst_v2026.8.17.1.json'>2026.8.17.1</a></td><td class='num'>0.393</td><td>improved</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization/hst:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/pixelization/hst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/pixelization_breakdown_hst_v2026.5.29.4_sparse.json'>2026.5.29.4</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/pixelization_breakdown_hst_v2026.7.6.649_sparse.json'>2026.7.6.649</a></td><td class='num'>1.139</td><td>flat</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization/hst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization/hst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>breakdown:imaging/pixelization_numba/euclid:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization_numba/euclid:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization_numba/euclid:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/pixelization_numba/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization_numba/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization_numba/hst:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_direct_conv/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_direct_conv/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_direct_conv/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_direct_conv/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_direct_conv/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_direct_conv/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_jax/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_jax/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_jax/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_jax/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_jax/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_jax/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_reference/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_reference/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_reference/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_reference/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_reference/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_reference/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_source_loop/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_source_loop/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_source_loop/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_symmetric/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_symmetric/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_symmetric/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_two_stage/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_two_stage/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_two_stage/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/mge/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/mge/alma:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge/alma:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge/alma:local_cpu_fp64_sparse</button></td></tr><tr><td><code>breakdown:interferometer/mge/alma_high:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge/alma_high:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge/alma_high:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/mge/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/mge/sma:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge/sma:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge/sma:local_cpu_fp64_sparse</button></td></tr><tr><td><code>breakdown:interferometer/mge_dft/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge_dft/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge_dft/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba_direct_conv/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba_direct_conv/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba_direct_conv/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba_direct_conv/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba_direct_conv/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba_direct_conv/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba_jax/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba_jax/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba_jax/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba_jax/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba_jax/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba_jax/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba_reference/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba_reference/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba_reference/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba_reference/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba_reference/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba_reference/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:point_source_image/image_plane/simple:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:point_source_image/image_plane/simple:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:point_source_image/image_plane/simple:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:point_source_source/source_plane/simple:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:point_source_source/source_plane/simple:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:point_source_source/source_plane/simple:local_cpu_fp64</button></td></tr></tbody></table></div><h3>Comparisons on cpu · cpu · mixed</h3><div class='tablewrap'><table><thead><tr><th>Comparison</th><th>Policy</th><th>Axis · metric</th><th>Baseline → candidate</th><th>Ratio (producer)</th><th>Status</th><th>Qualified</th><th>Reasons</th><th>Triage</th></tr></thead><tbody><tr><td><code>runtime:imaging/delaunay/hst:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/hst:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/hst:local_cpu_mp</button></td></tr><tr><td><code>runtime:imaging/delaunay/hst:local_cpu_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/hst:local_cpu_mp_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/hst:local_cpu_mp_sparse</button></td></tr><tr><td><code>runtime:imaging/delaunay/jwst:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/jwst:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/jwst:local_cpu_mp</button></td></tr><tr><td><code>runtime:imaging/delaunay/jwst:local_cpu_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/jwst:local_cpu_mp_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/jwst:local_cpu_mp_sparse</button></td></tr><tr><td><code>runtime:imaging/mge/ao:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/ao:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/ao:local_cpu_mp</button></td></tr><tr><td><code>runtime:imaging/mge/hst:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/hst:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/hst:local_cpu_mp</button></td></tr><tr><td><code>runtime:imaging/mge/hst:local_cpu_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/hst:local_cpu_mp_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/hst:local_cpu_mp_sparse</button></td></tr><tr><td><code>runtime:imaging/mge/jwst:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/jwst:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/jwst:local_cpu_mp</button></td></tr><tr><td><code>runtime:imaging/mge/jwst:local_cpu_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/jwst:local_cpu_mp_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/jwst:local_cpu_mp_sparse</button></td></tr><tr><td><code>runtime:imaging/pixelization/hst:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/hst:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/hst:local_cpu_mp</button></td></tr><tr><td><code>runtime:imaging/pixelization/hst:local_cpu_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/hst:local_cpu_mp_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/hst:local_cpu_mp_sparse</button></td></tr><tr><td><code>runtime:imaging/pixelization/jwst:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/jwst:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/jwst:local_cpu_mp</button></td></tr><tr><td><code>runtime:imaging/pixelization/jwst:local_cpu_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/jwst:local_cpu_mp_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/jwst:local_cpu_mp_sparse</button></td></tr><tr><td><code>runtime:interferometer/delaunay/alma:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:interferometer/delaunay/alma:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:interferometer/delaunay/alma:local_cpu_mp</button></td></tr><tr><td><code>runtime:interferometer/delaunay/sma:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:interferometer/delaunay/sma:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:interferometer/delaunay/sma:local_cpu_mp</button></td></tr><tr><td><code>runtime:interferometer/mge/sma:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:interferometer/mge/sma:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:interferometer/mge/sma:local_cpu_mp</button></td></tr><tr><td><code>runtime:interferometer/pixelization/sma:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:interferometer/pixelization/sma:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:interferometer/pixelization/sma:local_cpu_mp</button></td></tr></tbody></table></div><h3>Comparisons on a100 · gpu · float64</h3><div class='tablewrap'><table><thead><tr><th>Comparison</th><th>Policy</th><th>Axis · metric</th><th>Baseline → candidate</th><th>Ratio (producer)</th><th>Status</th><th>Qualified</th><th>Reasons</th><th>Triage</th></tr></thead><tbody><tr><td><code>runtime:imaging/delaunay/hst:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/hst:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/hst:hpc_a100_fp64</button></td></tr><tr><td><code>runtime:imaging/delaunay/hst:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/hst:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/hst:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/delaunay_nn/hst:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay_nn/hst:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay_nn/hst:hpc_a100_fp64</button></td></tr><tr><td><code>runtime:imaging/delaunay_nn/hst:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay_nn/hst:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay_nn/hst:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/pixelization/hst:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/hst:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/hst:hpc_a100_fp64</button></td></tr><tr><td><code>runtime:imaging/pixelization/hst:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/hst:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/hst:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>runtime:point_source_source/source_plane_solved:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:point_source_source/source_plane_solved:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage runtime:point_source_source/source_plane_solved:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:datacube/delaunay/alma_high:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:datacube/delaunay/alma_high:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:datacube/delaunay/alma_high:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:imaging/delaunay/hst:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:imaging/delaunay/hst:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>breakdown:imaging/delaunay_nn/hst:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay_nn/hst:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay_nn/hst:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:imaging/delaunay_nn/hst:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay_nn/hst:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay_nn/hst:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>breakdown:imaging/mge/hst:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/mge/hst:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/mge/hst:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:imaging/pixelization/hst:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:imaging/pixelization/hst:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/delaunay/alma_high:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/delaunay/alma_high:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/delaunay/alma_high:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/pixelization/alma_high:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/pixelization/alma_high:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/pixelization/alma_high:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay/alma:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay/alma:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay/alma:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/jvla/delaunay/jvla:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/jvla/delaunay/jvla:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/jvla/delaunay/jvla:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/jvla/mge/jvla:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/jvla/mge/jvla:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/jvla/mge/jvla:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/jvla/mge/jvla:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/jvla/mge/jvla:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/jvla/mge/jvla:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>breakdown:interferometer/jvla/pixelization/jvla:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/jvla/pixelization/jvla:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/jvla/pixelization/jvla:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/mge/alma:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge/alma:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge/alma:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/mge/alma:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge/alma:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge/alma:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>breakdown:interferometer/pixelization/alma:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization/alma:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization/alma:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/sma/delaunay/sma:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/delaunay/sma:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/delaunay/sma:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/sma/mge/sma:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/mge/sma:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/mge/sma:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/sma/pixelization/sma:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/pixelization/sma:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/pixelization/sma:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:point_source_source/source_plane/simple:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:point_source_source/source_plane/simple:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:point_source_source/source_plane/simple:hpc_a100_fp64</button></td></tr></tbody></table></div><h3>Comparisons on cpu · backend unknown · float64</h3><div class='tablewrap'><table><thead><tr><th>Comparison</th><th>Policy</th><th>Axis · metric</th><th>Baseline → candidate</th><th>Ratio (producer)</th><th>Status</th><th>Qualified</th><th>Reasons</th><th>Triage</th></tr></thead><tbody><tr><td><code>runtime:point_source_image/image_plane_solved:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:point_source_image/image_plane_solved:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:point_source_image/image_plane_solved:local_cpu_fp64</button></td></tr><tr><td><code>runtime:point_source_source/source_plane_solved:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:point_source_source/source_plane_solved:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:point_source_source/source_plane_solved:local_cpu_fp64</button></td></tr></tbody></table></div><h3>Comparisons on ral_cpu · cpu · float64</h3><div class='tablewrap'><table><thead><tr><th>Comparison</th><th>Policy</th><th>Axis · metric</th><th>Baseline → candidate</th><th>Ratio (producer)</th><th>Status</th><th>Qualified</th><th>Reasons</th><th>Triage</th></tr></thead><tbody><tr><td><code>runtime:point_source_source/source_plane_solved:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:point_source_source/source_plane_solved:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:point_source_source/source_plane_solved:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:cluster/image_plane:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:cluster/image_plane:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:cluster/image_plane:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/delaunay/alma_high:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/delaunay/alma_high:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/delaunay/alma_high:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/delaunay_numba/alma_high:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/delaunay_numba/alma_high:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/delaunay_numba/alma_high:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/pixelization/alma_high:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/pixelization/alma_high:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/pixelization/alma_high:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/pixelization_numba/alma_high:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/pixelization_numba/alma_high:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/pixelization_numba/alma_high:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay/alma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay/alma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay/alma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba/alma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba/alma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba/alma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization/alma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization/alma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization/alma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba/alma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba/alma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba/alma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/sma/delaunay/sma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/delaunay/sma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/delaunay/sma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/sma/delaunay_numba/sma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/delaunay_numba/sma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/delaunay_numba/sma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/sma/pixelization/sma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/pixelization/sma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/pixelization/sma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/sma/pixelization_numba/sma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/pixelization_numba/sma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/pixelization_numba/sma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:point_source_image/image_plane/simple:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:point_source_image/image_plane/simple:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:point_source_image/image_plane/simple:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:point_source_source/source_plane/simple:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:point_source_source/source_plane/simple:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:point_source_source/source_plane/simple:hpc_ral_cpu_fp64</button></td></tr></tbody></table></div><h3>Comparisons on a100 · gpu · mixed</h3><div class='tablewrap'><table><thead><tr><th>Comparison</th><th>Policy</th><th>Axis · metric</th><th>Baseline → candidate</th><th>Ratio (producer)</th><th>Status</th><th>Qualified</th><th>Reasons</th><th>Triage</th></tr></thead><tbody><tr><td><code>breakdown:datacube/delaunay/alma_high:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:datacube/delaunay/alma_high:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:datacube/delaunay/alma_high:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:imaging/delaunay/hst:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:imaging/delaunay/hst:hpc_a100_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_mp_sparse' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_mp_sparse</button></td></tr><tr><td><code>breakdown:imaging/pixelization/hst:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:imaging/pixelization/hst:hpc_a100_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_mp_sparse' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_mp_sparse</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/delaunay/alma_high:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/delaunay/alma_high:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/delaunay/alma_high:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/pixelization/alma_high:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/pixelization/alma_high:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/pixelization/alma_high:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/delaunay/alma:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay/alma:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay/alma:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/jvla/delaunay/jvla:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/jvla/delaunay/jvla:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/jvla/delaunay/jvla:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/jvla/mge/jvla:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/jvla/mge/jvla:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/jvla/mge/jvla:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/jvla/pixelization/jvla:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/jvla/pixelization/jvla:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/jvla/pixelization/jvla:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/pixelization/alma:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization/alma:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization/alma:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/sma/delaunay/sma:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/delaunay/sma:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/delaunay/sma:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/sma/pixelization/sma:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/pixelization/sma:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/pixelization/sma:hpc_a100_mp</button></td></tr></tbody></table></div></main><script>
+</style></head><body><header class='hero'><h1>PyAuto<span>Pulse</span></h1><p>Measure. Trace. Compare.</p></header><main><!-- pulse-campaigns:da22f6163319be7638016e9db58551fd0d6ad8164e45bd523af8174a4a913d7e --><section class="checkin"><h2>Check in on all profiling work</h2><p>One prompt, one ongoing chat. Add a campaign focus or idea here, or tell the chat before or afterwards.</p><label for="checkin-prompt">Your check-in prompt</label><textarea id="checkin-prompt" rows="8">Use PyAutoPulse as the home for all profiling work in this chat. Read PyAutoPulse/AGENTS.md and CHECKIN.md, then campaigns.yaml, the linked tasks and current project campaign ledgers. Check every open campaign and task for updates since the last check-in: results, jobs, PRs, releases, blockers and the next bounded step. Distinguish verified updates from stale or unavailable evidence; do not infer success from missing data. Give me one concise campaign-by-campaign summary and a proposed priority order. Apply any campaign direction or ideas I supply before or after this prompt; otherwise cover everything. Keep the overall check-in even when I focus one campaign. Update the Pulse ledger with verified facts and dated source links, regenerate the board, and use the repository workflow to persist changes. Use Brain&#x27;s profiling conductor for scientific judgement and project drivers for execution. Do not launch new compute, change defaults or baselines, or bypass a human/release gate merely to check in. Continue managing subsequent profiling requests in this chat. Optional direction: [leave blank, name a campaign, or suggest an idea].</textarea><button type="button" id="copy-checkin">Copy check-in prompt</button><span id="copy-status" role="status" aria-live="polite"></span><p class="muted">Last check-in: not recorded yet. Ledger dates are review dates, not measurement freshness.</p></section><h2 id="campaigns">Active campaigns</h2><div class="tablewrap"><table><thead><tr><th>Campaign</th><th>Status</th><th>Open tasks</th><th>Next step</th><th>Reviewed</th></tr></thead><tbody><tr><td><a href="https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/point_source_image_plane_cpu.md">Point-source image plane · CPU</a></td><td>active</td><td>1</td><td>Phases 1–4 shipped; reconcile carried cleanup and test-placement work.</td><td>2026-10-03</td></tr><tr><td><a href="https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/point_source_gpu_breakdown.md">Point-source image plane · A100</a></td><td>needs-decision</td><td>1</td><td>Measure an image-plane fit for the admission decision; retain forward-gradient correctness findings.</td><td>2026-10-03</td></tr><tr><td><a href="https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/point_source_source_plane.md">Point-source source plane</a></td><td>parked</td><td>2</td><td>Core complete; gradient-sampler admission required before blackjax work. Warm-up measurement fix remains open.</td><td>2026-10-03</td></tr><tr><td><a href="https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/cluster_pointsolver.md">Cluster PointSolver robustness and performance</a></td><td>active</td><td>2</td><td>Reconcile the expanded robustness contract; select one bounded phase and current representative baseline.</td><td>2026-10-03</td></tr><tr><td><a href="https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/interferometer_likelihood.md">Interferometer likelihood and streaming</a></td><td>active</td><td>5</td><td>Collect the final CPU matrix cell; scope streaming scaling, curvature reuse and FFT levers separately.</td><td>2026-10-03</td></tr><tr><td><a href="https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/linear_solver_accuracy.md">Linear-solver accuracy and cost</a></td><td>blocked</td><td>1</td><td>Verify release containing Array#595 and RAL sync before GPU/vmap parity and timing.</td><td>2026-10-03</td></tr><tr><td><a href="https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/post_certified_breakdown.md">Production imaging and likelihood breakdown</a></td><td>active</td><td>3</td><td>Reconcile measured GPU/CPU coverage, then profile production over-sampling and remaining costs.</td><td>2026-10-03</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoMind/tree/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft">Measurement reliability and profiling tools</a></td><td>active</td><td>9</td><td>Audit timing noise; repair warm-up, cumulative accounting, memo blindness and compile-cell construction.</td><td>2026-10-03</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoMind/tree/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft">PyAutoFit profiling bootstrap</a></td><td>needs-decision</td><td>1</td><td>Confirm new repository creation before porting baseline packages; no empty producer registration.</td><td>2026-10-03</td></tr><tr><td><a href="https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/critical_curves.md">Critical curves and evaluation grids</a></td><td>blocked</td><td>0</td><td>Galaxy#646 and profiling#365 merged; workspace#343 remains release-gated in Mind.</td><td>2026-10-03</td></tr><tr><td><a href="https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/certified_positive_solver.md">Certified positive solver</a></td><td>parked</td><td>0</td><td>Retain the parked opt-in solver study; read the project ledger before proposing new work.</td><td>2026-10-03</td></tr></tbody></table></div><h2 id="tasks">Active tasks</h2><p>Open means tracked, not necessarily running. Blockers and decisions still apply.</p><div class="tablewrap"><table><thead><tr><th>Task</th><th>Campaign</th><th>Status</th><th>Priority</th><th>Next step</th></tr></thead><tbody><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/runtime_cell_single_jit_gpu_warmup.md">Runtime cells&#x27; A100 `single_jit` includes the post-compile warm-up — source-plane 0.642 ms vs a steady 0.267 ms on the same node</a></td><td>point-source-source</td><td>needs-decision</td><td>low</td><td>Read the task and resolve its explicit human decision before execution.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/timing_noise_audit.md">Audit timing tests and profiling gates for measurement noise</a> · <a href="https://github.com/PyAutoLabs/autolens_profiling/issues/362">issue</a></td><td>measurement-tools</td><td>ready</td><td>high</td><td>Read the task and select the next bounded step; check current repository claims.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/mge_likelihood_breakdown_steps_are_cumulative_an.md">MGE likelihood_breakdown steps are cumulative and `linear_gaussians` is reported as 0</a></td><td>measurement-tools</td><td>ready</td><td>low</td><td>Read the task and select the next bounded step; check current repository claims.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/gradient_cost_probe.md">A gradient-cost probe: forward vs `value_and_grad` ms/eval and a strict FD check, on any registry cell</a></td><td>measurement-tools</td><td>ready</td><td>low</td><td>Read the task and select the next bounded step; check current repository claims.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/numba_breakdown_harness_memo_blind.md">Numba breakdown harness: perturb the instance so the operated-matrix memo cannot hide a step</a></td><td>measurement-tools</td><td>ready</td><td>medium</td><td>Read the task and select the next bounded step; check current repository claims.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/search_settings_estimation_infrastructure.md">Search settings-estimation + profiling infrastructure (n_starts / batch_size / n_batch)</a></td><td>measurement-tools</td><td>ready</td><td>normal</td><td>Read the task and select the next bounded step; check current repository claims.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/jax_compile_probe_needs_own_cell_builder.md">jax_compile/probe.py lost its cell builder with the searches tier — give profiling its own</a></td><td>measurement-tools</td><td>ready</td><td>normal</td><td>Read the task and select the next bounded step; check current repository claims.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/mge_nnls_fix_pyautoarray_571_slam_60.md">Linear-solver programme phase 3: GPU/vmap/A100 timing and parity rows for the solver corpus</a></td><td>linear-solver</td><td>blocked</td><td>medium</td><td>a release shipping the phase-2 fix (PyAutoArray#595, merged 2026-09-30 pending release; record complete/2026/09/raw-pdip-forward-polish.md) AND the RAL mirror synced via HPCPullPyAuto</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/autofit_profiling_bootstrap.md">autofit_profiling: bootstrap the repo + general PyAutoFit profiling epic</a></td><td>autofit</td><td>needs-decision</td><td>normal</td><td>Read the task and resolve its explicit human decision before execution.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/cluster_pointsolver_speed.md">Cluster PointSolver — robustness, analysis settings and performance</a></td><td>cluster-pointsolver</td><td>ready</td><td>normal</td><td>Read the task and select the next bounded step; check current repository claims.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/imaging_production_over_sampling.md">Profile imaging pixelizations at production over-sampling</a></td><td>imaging</td><td>ready</td><td>normal</td><td>Read the task and select the next bounded step; check current repository claims.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_decision_matrix_last_cell.md">Interferometer decision matrix — fill the last cell (CPU rect 39² at alma_high r5.0) and flip the wiki row to shipped</a></td><td>interferometer</td><td>needs-decision</td><td>normal</td><td>Read the task and resolve its explicit human decision before execution.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_fixed_mapper_curvature_preload.md">Interferometer fixed-mapper searches: reuse the W~ curvature matrix across likelihood calls via `preloads.curvature_matrix`</a></td><td>interferometer</td><td>needs-slicing</td><td>normal</td><td>Scope one bounded phase from the full task contract.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_nnls_memo_scattered_stream_guard.md">fnnls warm-start memo: stop it slowing scattered evaluation streams (interferometer CPU, autolens_profiling#332)</a></td><td>interferometer</td><td>ready</td><td>medium</td><td>Read the task and select the next bounded step; check current repository claims.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_streaming_scaling.md">Campaign: interferometer streaming (array-free) vs in-memory — memory and time scaling to 2e8 visibilities</a></td><td>interferometer</td><td>ready</td><td>high</td><td>Read the task and select the next bounded step; check current repository claims.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_w_tilde_fft_size_levers.md">Interferometer W~ curvature matrix is FFT-bound on the mask extent: pruned padded FFT and real-space pixel scale on the A100</a></td><td>interferometer</td><td>ready</td><td>normal</td><td>Read the task and select the next bounded step; check current repository claims.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/point_solver_profiling_cells.md">PointSolver profiling cells: lensed quasar → cluster runtime tier → single/multi-source → multiplane</a></td><td>cluster-pointsolver</td><td>ready</td><td>normal</td><td>Read the task and select the next bounded step; check current repository claims.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/point_source_image_plane_gpu_breakdown.md">Point-source A100 speed-up campaign: profile and optimize with the shared breakdown</a></td><td>point-source-image-gpu</td><td>needs-decision</td><td>high</td><td>Read the task and resolve its explicit human decision before execution.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/point_source_source_plane_chi_squared_speed.md">Point-source source-plane chi-squared speed-up campaign — remaining candidates (phases 1–2e shipped)</a></td><td>point-source-source</td><td>blocked</td><td>low</td><td>Await gradient-sampler admission evidence; core campaign already complete.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/pointsolver_cpu_speed_campaign_remainder.md">Point-source (single-source) CPU campaign — carried leftovers and completion evidence</a></td><td>point-source-image-cpu</td><td>ready</td><td>low</td><td>Read the task and select the next bounded step; check current repository claims.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/post_certified_solver_likelihood_breakdown.md">Assess remaining likelihood bottlenecks after certified solver integration on CPU and GPUs</a></td><td>imaging</td><td>needs-slicing</td><td>normal</td><td>Scope one bounded phase from the full task contract.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/profile_lens_aggregator_needs_config_dir.md">profile_lens_aggregator.py cannot run from the autolens_workspace_developer root: no config/ directory</a></td><td>measurement-tools</td><td>ready</td><td>low</td><td>Read the task, reconcile current evidence and select one bounded step.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/gradient_pixelization_pin_residual_drift.md">jax_profiling/gradient/imaging/pixelization.py: 3.2% of its pin move is unattributed</a></td><td>measurement-tools</td><td>ready</td><td>normal</td><td>Read the task, reconcile current evidence and select one bounded step.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/pair_jax_xla_env_vars_with_measured.md">Pair JAX/XLA env vars with measured compile and run times, per backend</a></td><td>measurement-tools</td><td>ready</td><td>high</td><td>Read the task, reconcile current evidence and select one bounded step.</td></tr><tr><td><a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/quick_update_plotting_cost.md">Quick-update plotting cost — minutes per update, and it is not JAX compile</a></td><td>imaging</td><td>ready</td><td>medium</td><td>Read the task, reconcile current evidence and select one bounded step.</td></tr></tbody></table></div><p>Completed and superseded tasks remain in the <a href="https://github.com/PyAutoLabs/PyAutoPulse/blob/main/campaigns.yaml">ledger</a>.</p><h2 id="evidence">Profiling evidence</h2><p class='lede'>The cross-project view of the organism&#x27;s profiling. Each <code>&lt;lib&gt;_profiling</code> project repo runs its producers, keeps its results and drift policy, and publishes a <code>profiling-summary</code> file. This page reads each file at one resolved commit, validates the exchange contract and shows what the producer says: it recomputes no ratio, combines no timings whose axis or hardware differs, ranks nothing across projects and issues no verdict. Judgment belongs to the Brain&#x27;s profiling conductor (<code>/profiling triage</code>).</p><div class='stats'><div><b>1</b><span>Projects</span></div><div><b>159</b><span>Records</span></div><div><b>145</b><span>Comparisons</span></div><div><b>0</b><span>Drifted</span></div><div><b>0</b><span>Refused pairs</span></div><div><b>0</b><span>Cached</span></div><div><b>0</b><span>Failed</span></div></div><h2 id='projects'>Projects</h2><div class='tablewrap'><table><thead><tr><th>Project</th><th>Scope</th><th>Evidence</th><th>Last fetch</th><th>Coverage</th><th>Integrity</th><th>Freshness</th><th>Qualification</th><th>Links</th></tr></thead><tbody><tr><td><a href='#lens'>lens</a> <span class='muted'>autolens_profiling</span></td><td>release-runtime</td><td>2026-08-17</td><td>2026-10-03 @ efcece3e</td><td>145 series · 159 points · 4 releases · 61 cells · 6 excluded</td><td class='ok'>ok</td><td>freshness policy unspecified · age 47 d at fetch 2026-10-03</td><td>2/159 records · 0/145 comparisons qualified</td><td><a href='https://pyautolabs.github.io/autolens_profiling/'>dashboard</a> · <a href='https://github.com/PyAutoLabs/autolens_profiling'>repo</a> · <a href='https://github.com/PyAutoLabs/PyAutoPulse/blob/main/receipts/lens.json'>receipt</a></td></tr></tbody></table></div><p class='muted'>Three notions are kept apart: <b>integrity</b> is whether the file was fetched and meets the contract; <b>freshness</b> is the evidence&#x27;s age against the producer&#x27;s own <code>valid_until</code> (none declared → &quot;freshness policy unspecified&quot;); <b>qualification</b> is the producer&#x27;s own flag on each record and comparison. Healthy transport never makes unknown quality green.</p><h2 id='lens'>lens · autolens_profiling</h2><p class='lede'>Project <code>autolens_profiling</code>, scope <code>release-runtime</code>, read from <a href='https://github.com/PyAutoLabs/autolens_profiling'>PyAutoLabs/autolens_profiling</a> <code>dashboard/summary.json</code> at <a href='https://github.com/PyAutoLabs/autolens_profiling/tree/efcece3ea39c622e62cc61518f58394ee3986d4c'><code>efcece3e</code></a>; producer revision <code>227b5c91</code>; generated 2026-09-30T19:43:00Z; policy <code>runtime-drift-2x-1ms</code>. <a href='https://pyautolabs.github.io/autolens_profiling/'>Project dashboard</a> · <a href='https://github.com/PyAutoLabs/PyAutoPulse/blob/main/receipts/lens.json'>receipt</a>.</p><div class='notions'><div><b>Integrity</b><span class='ok'>ok</span></div><div><b>Freshness</b>freshness policy unspecified · age 47 d at fetch 2026-10-03 (evidence 2026-08-17)<span class='muted' data-age-from='2026-08-17T00:00:00Z'></span></div><div><b>Qualification</b>2/159 records · 0/145 comparisons qualified</div></div><h3>Limitations (the producer's own)</h3><ul><li>runtime axis only: compile times, VRAM and per-step component timings are not in this feed</li><li>evidence_updated_at is the newest release date encoded in a measured library version; result rows do not record measurement wall-clock time</li><li>valid_until is null: this project declares no freshness policy for release trends</li><li>vmap_per_call_s is null where a row carries no vmap block</li><li>rows without a device.provenance block (pre-autolens_profiling#342) are unqualified</li><li>6 row(s) above the load-average cap 8 are listed under coverage.excluded and never enter records</li></ul><h3>Excluded by the producer</h3><ul><li><code>breakdown:interferometer/sdp81/delaunay/sdp81:hpc_a100_fp64@2026.8.17.1</code>: loadavg 65.7 above the cap 8 (<a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/interferometer/sdp81/delaunay_hpc_a100_fp64.json'>evidence</a>)</li><li><code>breakdown:interferometer/sdp81/delaunay_numba/sdp81:hpc_ral_cpu_fp64@2026.8.17.1</code>: loadavg 68.3 above the cap 8 (<a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/interferometer/sdp81/delaunay_numba_hpc_ral_cpu_fp64.json'>evidence</a>)</li><li><code>breakdown:interferometer/sdp81/mge/sdp81:hpc_a100_fp64@2026.8.17.1</code>: loadavg 55.7 above the cap 8 (<a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/interferometer/sdp81/mge_hpc_a100_fp64.json'>evidence</a>)</li><li><code>breakdown:interferometer/sdp81/mge/sdp81:hpc_a100_fp64_sparse@2026.8.17.1</code>: loadavg 55.7 above the cap 8 (<a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/interferometer/sdp81/mge_hpc_a100_fp64_sparse.json'>evidence</a>)</li><li><code>breakdown:interferometer/sdp81/pixelization/sdp81:hpc_a100_fp64@2026.8.17.1</code>: loadavg 65.7 above the cap 8 (<a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/interferometer/sdp81/pixelization_hpc_a100_fp64.json'>evidence</a>)</li><li><code>breakdown:interferometer/sdp81/pixelization_numba/sdp81:hpc_ral_cpu_fp64@2026.8.17.1</code>: loadavg 68.3 above the cap 8 (<a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/interferometer/sdp81/pixelization_numba_hpc_ral_cpu_fp64.json'>evidence</a>)</li></ul><h3>Comparisons on cpu · cpu · float64</h3><div class='tablewrap'><table><thead><tr><th>Comparison</th><th>Policy</th><th>Axis · metric</th><th>Baseline → candidate</th><th>Ratio (producer)</th><th>Status</th><th>Qualified</th><th>Reasons</th><th>Triage</th></tr></thead><tbody><tr><td><code>runtime:imaging/delaunay/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/runtime/imaging/delaunay/hst/comparison.json'>2026.7.6.649 [local_cpu_fp64]</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/runtime/imaging/delaunay/delaunay_likelihood_summary_hst_v2026.8.17.1.json'>2026.8.17.1</a></td><td class='num'>0.27</td><td>improved</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/delaunay/hst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/hst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/hst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/delaunay/jwst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/jwst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/jwst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/delaunay/jwst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/jwst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/jwst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/delaunay_nn/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay_nn/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay_nn/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/delaunay_numba/euclid:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay_numba/euclid:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay_numba/euclid:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/delaunay_numba/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay_numba/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay_numba/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/mge/ao:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/ao:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/ao:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/mge/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/runtime/imaging/mge/hst/comparison.json'>2026.7.6.649 [local_cpu_fp64]</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/runtime/imaging/mge/mge_likelihood_summary_hst_v2026.8.17.1.json'>2026.8.17.1</a></td><td class='num'>0.671</td><td>flat</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/mge/hst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/hst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/hst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/mge/jwst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/jwst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/jwst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/mge/jwst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/jwst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/jwst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/mge_mass_jax/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge_mass_jax/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge_mass_jax/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/pixelization/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/runtime/imaging/pixelization/hst/comparison.json'>2026.7.6.649 [local_cpu_fp64]</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/runtime/imaging/pixelization/pixelization_likelihood_summary_hst_v2026.8.17.1.json'>2026.8.17.1</a></td><td class='num'>0.232</td><td>improved</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/pixelization/hst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/hst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/hst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/pixelization/jwst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/jwst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/jwst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/pixelization/jwst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/jwst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/jwst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/pixelization_numba/euclid:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization_numba/euclid:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization_numba/euclid:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/pixelization_numba/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization_numba/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization_numba/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:imaging/pixelization_numba_mge_mass/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization_numba_mge_mass/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization_numba_mge_mass/hst:local_cpu_fp64</button></td></tr><tr><td><code>runtime:interferometer/delaunay/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:interferometer/delaunay/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:interferometer/delaunay/sma:local_cpu_fp64</button></td></tr><tr><td><code>runtime:interferometer/mge/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:interferometer/mge/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:interferometer/mge/sma:local_cpu_fp64</button></td></tr><tr><td><code>runtime:interferometer/pixelization/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:interferometer/pixelization/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:interferometer/pixelization/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:cluster/image_plane:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/cluster/image_plane_breakdown_v2026.7.23.1.json'>2026.7.23.1</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/cluster/image_plane_local_cpu_fp64.json'>2026.8.17.1</a></td><td class='num'>0.469</td><td>improved</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage breakdown:cluster/image_plane:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:cluster/image_plane:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:cluster/source_plane:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/cluster/source_plane_breakdown_v2026.7.6.649.json'>2026.7.6.649</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/cluster/source_plane_breakdown_v2026.7.23.1.json'>2026.7.23.1</a></td><td class='num'>1.469</td><td>flat</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage breakdown:cluster/source_plane:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:cluster/source_plane:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/delaunay/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/delaunay_breakdown_hst_v2026.7.6.649.json'>2026.7.6.649</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/delaunay_breakdown_hst_v2026.8.17.1.json'>2026.8.17.1</a></td><td class='num'>0.331</td><td>improved</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay/hst:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/delaunay/hst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/delaunay_breakdown_hst_v2026.5.29.4_sparse.json'>2026.5.29.4</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/delaunay_breakdown_hst_v2026.7.6.649_sparse.json'>2026.7.6.649</a></td><td class='num'>1.108</td><td>flat</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay/hst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay/hst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>breakdown:imaging/delaunay_nn/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay_nn/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay_nn/hst:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/delaunay_numba/euclid:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay_numba/euclid:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay_numba/euclid:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/delaunay_numba/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay_numba/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay_numba/hst:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/mge/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/mge_breakdown_hst_v2026.7.6.649.json'>2026.7.6.649</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/mge_breakdown_hst_v2026.8.17.1.json'>2026.8.17.1</a></td><td class='num'>0.225</td><td>improved</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/mge/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/mge/hst:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/pixelization/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/pixelization_breakdown_hst_v2026.7.6.649.json'>2026.7.6.649</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/pixelization_breakdown_hst_v2026.8.17.1.json'>2026.8.17.1</a></td><td class='num'>0.393</td><td>improved</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization/hst:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/pixelization/hst:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td><a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/pixelization_breakdown_hst_v2026.5.29.4_sparse.json'>2026.5.29.4</a> → <a href='https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/results/breakdown/imaging/pixelization_breakdown_hst_v2026.7.6.649_sparse.json'>2026.7.6.649</a></td><td class='num'>1.139</td><td>flat</td><td>no</td><td>baseline row unqualified: no provenance block (pre-#342 row); candidate row unqualified: no provenance block (pre-#342 row)</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization/hst:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization/hst:local_cpu_fp64_sparse</button></td></tr><tr><td><code>breakdown:imaging/pixelization_numba/euclid:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization_numba/euclid:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization_numba/euclid:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:imaging/pixelization_numba/hst:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization_numba/hst:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization_numba/hst:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_direct_conv/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_direct_conv/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_direct_conv/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_direct_conv/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_direct_conv/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_direct_conv/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_jax/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_jax/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_jax/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_jax/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_jax/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_jax/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_reference/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_reference/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_reference/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_reference/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_reference/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_reference/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_source_loop/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_source_loop/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_source_loop/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_symmetric/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_symmetric/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_symmetric/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba_two_stage/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba_two_stage/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba_two_stage/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/mge/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/mge/alma:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge/alma:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge/alma:local_cpu_fp64_sparse</button></td></tr><tr><td><code>breakdown:interferometer/mge/alma_high:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge/alma_high:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge/alma_high:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/mge/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/mge/sma:local_cpu_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge/sma:local_cpu_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge/sma:local_cpu_fp64_sparse</button></td></tr><tr><td><code>breakdown:interferometer/mge_dft/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge_dft/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge_dft/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba_direct_conv/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba_direct_conv/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba_direct_conv/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba_direct_conv/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba_direct_conv/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba_direct_conv/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba_jax/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba_jax/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba_jax/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba_jax/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba_jax/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba_jax/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba_reference/alma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba_reference/alma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba_reference/alma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba_reference/sma:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba_reference/sma:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba_reference/sma:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:point_source_image/image_plane/simple:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:point_source_image/image_plane/simple:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:point_source_image/image_plane/simple:local_cpu_fp64</button></td></tr><tr><td><code>breakdown:point_source_source/source_plane/simple:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:point_source_source/source_plane/simple:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:point_source_source/source_plane/simple:local_cpu_fp64</button></td></tr></tbody></table></div><h3>Comparisons on cpu · cpu · mixed</h3><div class='tablewrap'><table><thead><tr><th>Comparison</th><th>Policy</th><th>Axis · metric</th><th>Baseline → candidate</th><th>Ratio (producer)</th><th>Status</th><th>Qualified</th><th>Reasons</th><th>Triage</th></tr></thead><tbody><tr><td><code>runtime:imaging/delaunay/hst:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/hst:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/hst:local_cpu_mp</button></td></tr><tr><td><code>runtime:imaging/delaunay/hst:local_cpu_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/hst:local_cpu_mp_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/hst:local_cpu_mp_sparse</button></td></tr><tr><td><code>runtime:imaging/delaunay/jwst:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/jwst:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/jwst:local_cpu_mp</button></td></tr><tr><td><code>runtime:imaging/delaunay/jwst:local_cpu_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/jwst:local_cpu_mp_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/jwst:local_cpu_mp_sparse</button></td></tr><tr><td><code>runtime:imaging/mge/ao:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/ao:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/ao:local_cpu_mp</button></td></tr><tr><td><code>runtime:imaging/mge/hst:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/hst:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/hst:local_cpu_mp</button></td></tr><tr><td><code>runtime:imaging/mge/hst:local_cpu_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/hst:local_cpu_mp_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/hst:local_cpu_mp_sparse</button></td></tr><tr><td><code>runtime:imaging/mge/jwst:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/jwst:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/jwst:local_cpu_mp</button></td></tr><tr><td><code>runtime:imaging/mge/jwst:local_cpu_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/mge/jwst:local_cpu_mp_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/mge/jwst:local_cpu_mp_sparse</button></td></tr><tr><td><code>runtime:imaging/pixelization/hst:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/hst:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/hst:local_cpu_mp</button></td></tr><tr><td><code>runtime:imaging/pixelization/hst:local_cpu_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/hst:local_cpu_mp_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/hst:local_cpu_mp_sparse</button></td></tr><tr><td><code>runtime:imaging/pixelization/jwst:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/jwst:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/jwst:local_cpu_mp</button></td></tr><tr><td><code>runtime:imaging/pixelization/jwst:local_cpu_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/jwst:local_cpu_mp_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/jwst:local_cpu_mp_sparse</button></td></tr><tr><td><code>runtime:interferometer/delaunay/alma:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:interferometer/delaunay/alma:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:interferometer/delaunay/alma:local_cpu_mp</button></td></tr><tr><td><code>runtime:interferometer/delaunay/sma:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:interferometer/delaunay/sma:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:interferometer/delaunay/sma:local_cpu_mp</button></td></tr><tr><td><code>runtime:interferometer/mge/sma:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:interferometer/mge/sma:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:interferometer/mge/sma:local_cpu_mp</button></td></tr><tr><td><code>runtime:interferometer/pixelization/sma:local_cpu_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:interferometer/pixelization/sma:local_cpu_mp' title='Copy the manual handoff'>/profiling triage runtime:interferometer/pixelization/sma:local_cpu_mp</button></td></tr></tbody></table></div><h3>Comparisons on a100 · gpu · float64</h3><div class='tablewrap'><table><thead><tr><th>Comparison</th><th>Policy</th><th>Axis · metric</th><th>Baseline → candidate</th><th>Ratio (producer)</th><th>Status</th><th>Qualified</th><th>Reasons</th><th>Triage</th></tr></thead><tbody><tr><td><code>runtime:imaging/delaunay/hst:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/hst:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/hst:hpc_a100_fp64</button></td></tr><tr><td><code>runtime:imaging/delaunay/hst:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay/hst:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay/hst:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/delaunay_nn/hst:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay_nn/hst:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay_nn/hst:hpc_a100_fp64</button></td></tr><tr><td><code>runtime:imaging/delaunay_nn/hst:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/delaunay_nn/hst:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/delaunay_nn/hst:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>runtime:imaging/pixelization/hst:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/hst:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/hst:hpc_a100_fp64</button></td></tr><tr><td><code>runtime:imaging/pixelization/hst:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:imaging/pixelization/hst:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage runtime:imaging/pixelization/hst:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>runtime:point_source_source/source_plane_solved:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:point_source_source/source_plane_solved:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage runtime:point_source_source/source_plane_solved:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:datacube/delaunay/alma_high:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:datacube/delaunay/alma_high:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:datacube/delaunay/alma_high:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:imaging/delaunay/hst:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:imaging/delaunay/hst:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>breakdown:imaging/delaunay_nn/hst:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay_nn/hst:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay_nn/hst:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:imaging/delaunay_nn/hst:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay_nn/hst:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay_nn/hst:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>breakdown:imaging/mge/hst:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/mge/hst:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/mge/hst:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:imaging/pixelization/hst:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:imaging/pixelization/hst:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/delaunay/alma_high:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/delaunay/alma_high:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/delaunay/alma_high:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/pixelization/alma_high:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/pixelization/alma_high:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/pixelization/alma_high:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay/alma:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay/alma:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay/alma:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/jvla/delaunay/jvla:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/jvla/delaunay/jvla:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/jvla/delaunay/jvla:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/jvla/mge/jvla:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/jvla/mge/jvla:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/jvla/mge/jvla:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/jvla/mge/jvla:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/jvla/mge/jvla:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/jvla/mge/jvla:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>breakdown:interferometer/jvla/pixelization/jvla:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/jvla/pixelization/jvla:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/jvla/pixelization/jvla:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/mge/alma:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge/alma:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge/alma:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/mge/alma:hpc_a100_fp64_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/mge/alma:hpc_a100_fp64_sparse' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/mge/alma:hpc_a100_fp64_sparse</button></td></tr><tr><td><code>breakdown:interferometer/pixelization/alma:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization/alma:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization/alma:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/sma/delaunay/sma:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/delaunay/sma:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/delaunay/sma:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/sma/mge/sma:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/mge/sma:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/mge/sma:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:interferometer/sma/pixelization/sma:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/pixelization/sma:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/pixelization/sma:hpc_a100_fp64</button></td></tr><tr><td><code>breakdown:point_source_source/source_plane/simple:hpc_a100_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:point_source_source/source_plane/simple:hpc_a100_fp64' title='Copy the manual handoff'>/profiling triage breakdown:point_source_source/source_plane/simple:hpc_a100_fp64</button></td></tr></tbody></table></div><h3>Comparisons on cpu · backend unknown · float64</h3><div class='tablewrap'><table><thead><tr><th>Comparison</th><th>Policy</th><th>Axis · metric</th><th>Baseline → candidate</th><th>Ratio (producer)</th><th>Status</th><th>Qualified</th><th>Reasons</th><th>Triage</th></tr></thead><tbody><tr><td><code>runtime:point_source_image/image_plane_solved:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:point_source_image/image_plane_solved:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:point_source_image/image_plane_solved:local_cpu_fp64</button></td></tr><tr><td><code>runtime:point_source_source/source_plane_solved:local_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:point_source_source/source_plane_solved:local_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:point_source_source/source_plane_solved:local_cpu_fp64</button></td></tr></tbody></table></div><h3>Comparisons on ral_cpu · cpu · float64</h3><div class='tablewrap'><table><thead><tr><th>Comparison</th><th>Policy</th><th>Axis · metric</th><th>Baseline → candidate</th><th>Ratio (producer)</th><th>Status</th><th>Qualified</th><th>Reasons</th><th>Triage</th></tr></thead><tbody><tr><td><code>runtime:point_source_source/source_plane_solved:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage runtime:point_source_source/source_plane_solved:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage runtime:point_source_source/source_plane_solved:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:cluster/image_plane:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:cluster/image_plane:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:cluster/image_plane:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/delaunay/alma_high:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/delaunay/alma_high:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/delaunay/alma_high:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/delaunay_numba/alma_high:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/delaunay_numba/alma_high:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/delaunay_numba/alma_high:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/pixelization/alma_high:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/pixelization/alma_high:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/pixelization/alma_high:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/pixelization_numba/alma_high:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/pixelization_numba/alma_high:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/pixelization_numba/alma_high:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay/alma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay/alma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay/alma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/delaunay_numba/alma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay_numba/alma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay_numba/alma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization/alma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization/alma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization/alma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/pixelization_numba/alma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization_numba/alma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization_numba/alma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/sma/delaunay/sma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/delaunay/sma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/delaunay/sma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/sma/delaunay_numba/sma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/delaunay_numba/sma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/delaunay_numba/sma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/sma/pixelization/sma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/pixelization/sma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/pixelization/sma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:interferometer/sma/pixelization_numba/sma:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/pixelization_numba/sma:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/pixelization_numba/sma:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:point_source_image/image_plane/simple:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:point_source_image/image_plane/simple:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:point_source_image/image_plane/simple:hpc_ral_cpu_fp64</button></td></tr><tr><td><code>breakdown:point_source_source/source_plane/simple:hpc_ral_cpu_fp64</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:point_source_source/source_plane/simple:hpc_ral_cpu_fp64' title='Copy the manual handoff'>/profiling triage breakdown:point_source_source/source_plane/simple:hpc_ral_cpu_fp64</button></td></tr></tbody></table></div><h3>Comparisons on a100 · gpu · mixed</h3><div class='tablewrap'><table><thead><tr><th>Comparison</th><th>Policy</th><th>Axis · metric</th><th>Baseline → candidate</th><th>Ratio (producer)</th><th>Status</th><th>Qualified</th><th>Reasons</th><th>Triage</th></tr></thead><tbody><tr><td><code>breakdown:datacube/delaunay/alma_high:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:datacube/delaunay/alma_high:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:datacube/delaunay/alma_high:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:imaging/delaunay/hst:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:imaging/delaunay/hst:hpc_a100_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_mp_sparse' title='Copy the manual handoff'>/profiling triage breakdown:imaging/delaunay/hst:hpc_a100_mp_sparse</button></td></tr><tr><td><code>breakdown:imaging/pixelization/hst:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:imaging/pixelization/hst:hpc_a100_mp_sparse</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_mp_sparse' title='Copy the manual handoff'>/profiling triage breakdown:imaging/pixelization/hst:hpc_a100_mp_sparse</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/delaunay/alma_high:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/delaunay/alma_high:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/delaunay/alma_high:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/mge/alma_high:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/alma_high/pixelization/alma_high:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/alma_high/pixelization/alma_high:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/alma_high/pixelization/alma_high:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/delaunay/alma:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/delaunay/alma:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/delaunay/alma:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/jvla/delaunay/jvla:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/jvla/delaunay/jvla:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/jvla/delaunay/jvla:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/jvla/mge/jvla:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/jvla/mge/jvla:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/jvla/mge/jvla:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/jvla/pixelization/jvla:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/jvla/pixelization/jvla:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/jvla/pixelization/jvla:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/pixelization/alma:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/pixelization/alma:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/pixelization/alma:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/sma/delaunay/sma:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/delaunay/sma:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/delaunay/sma:hpc_a100_mp</button></td></tr><tr><td><code>breakdown:interferometer/sma/pixelization/sma:hpc_a100_mp</code></td><td>runtime-drift-2x-1ms</td><td>runtime · single_jit_s</td><td>— → —</td><td class='num'>—</td><td>insufficient</td><td>no</td><td>one release only</td><td><button type='button' data-copy='/profiling triage breakdown:interferometer/sma/pixelization/sma:hpc_a100_mp' title='Copy the manual handoff'>/profiling triage breakdown:interferometer/sma/pixelization/sma:hpc_a100_mp</button></td></tr></tbody></table></div></main><script>
+var copyCheckin=document.getElementById('copy-checkin');
+if(copyCheckin){copyCheckin.addEventListener('click',async function(){
+  var field=document.getElementById('checkin-prompt'), status=document.getElementById('copy-status');
+  try {
+    if(navigator.clipboard && window.isSecureContext){await navigator.clipboard.writeText(field.value)}
+    else {field.focus();field.select();if(!document.execCommand('copy')){throw new Error('copy unavailable')}}
+    status.textContent='Copied';
+  } catch(error){field.focus();field.select();status.textContent='Select and copy the prompt above.'}
+});}
+
 document.querySelectorAll('button[data-copy]').forEach(function(b){
   b.addEventListener('click',function(){
     var t=b.getAttribute('data-copy');
diff --git a/dashboard.md b/dashboard.md
index f6f6ce8..138fd7c 100644
--- a/dashboard.md
+++ b/dashboard.md
@@ -2,6 +2,69 @@
 
 <!-- Generated by `bin/pyauto-pulse board` from registry.yaml, receipts/ and snapshots/. Do not edit by hand. -->
 
+<!-- pulse-campaigns:da22f6163319be7638016e9db58551fd0d6ad8164e45bd523af8174a4a913d7e -->
+## Check in on all profiling work
+
+Copy this into one chat. Add a campaign focus or idea before or after it, or leave it unchanged.
+
+```text
+Use PyAutoPulse as the home for all profiling work in this chat. Read PyAutoPulse/AGENTS.md and CHECKIN.md, then campaigns.yaml, the linked tasks and current project campaign ledgers. Check every open campaign and task for updates since the last check-in: results, jobs, PRs, releases, blockers and the next bounded step. Distinguish verified updates from stale or unavailable evidence; do not infer success from missing data. Give me one concise campaign-by-campaign summary and a proposed priority order. Apply any campaign direction or ideas I supply before or after this prompt; otherwise cover everything. Keep the overall check-in even when I focus one campaign. Update the Pulse ledger with verified facts and dated source links, regenerate the board, and use the repository workflow to persist changes. Use Brain's profiling conductor for scientific judgement and project drivers for execution. Do not launch new compute, change defaults or baselines, or bypass a human/release gate merely to check in. Continue managing subsequent profiling requests in this chat. Optional direction: [leave blank, name a campaign, or suggest an idea].
+```
+
+Last check-in: not recorded yet. Ledger dates are review dates, not measurement freshness.
+
+## Active campaigns
+
+| Campaign | Status | Open tasks | Next step | Reviewed |
+|---|---|---:|---|---|
+| [Point-source image plane · CPU](https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/point_source_image_plane_cpu.md) | active | 1 | Phases 1–4 shipped; reconcile carried cleanup and test-placement work. | 2026-10-03 |
+| [Point-source image plane · A100](https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/point_source_gpu_breakdown.md) | needs-decision | 1 | Measure an image-plane fit for the admission decision; retain forward-gradient correctness findings. | 2026-10-03 |
+| [Point-source source plane](https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/point_source_source_plane.md) | parked | 2 | Core complete; gradient-sampler admission required before blackjax work. Warm-up measurement fix remains open. | 2026-10-03 |
+| [Cluster PointSolver robustness and performance](https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/cluster_pointsolver.md) | active | 2 | Reconcile the expanded robustness contract; select one bounded phase and current representative baseline. | 2026-10-03 |
+| [Interferometer likelihood and streaming](https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/interferometer_likelihood.md) | active | 5 | Collect the final CPU matrix cell; scope streaming scaling, curvature reuse and FFT levers separately. | 2026-10-03 |
+| [Linear-solver accuracy and cost](https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/linear_solver_accuracy.md) | blocked | 1 | Verify release containing Array#595 and RAL sync before GPU/vmap parity and timing. | 2026-10-03 |
+| [Production imaging and likelihood breakdown](https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/post_certified_breakdown.md) | active | 3 | Reconcile measured GPU/CPU coverage, then profile production over-sampling and remaining costs. | 2026-10-03 |
+| [Measurement reliability and profiling tools](https://github.com/PyAutoLabs/PyAutoMind/tree/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft) | active | 9 | Audit timing noise; repair warm-up, cumulative accounting, memo blindness and compile-cell construction. | 2026-10-03 |
+| [PyAutoFit profiling bootstrap](https://github.com/PyAutoLabs/PyAutoMind/tree/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft) | needs-decision | 1 | Confirm new repository creation before porting baseline packages; no empty producer registration. | 2026-10-03 |
+| [Critical curves and evaluation grids](https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/critical_curves.md) | blocked | 0 | Galaxy#646 and profiling#365 merged; workspace#343 remains release-gated in Mind. | 2026-10-03 |
+| [Certified positive solver](https://github.com/PyAutoLabs/autolens_profiling/blob/efcece3ea39c622e62cc61518f58394ee3986d4c/wiki/campaigns/certified_positive_solver.md) | parked | 0 | Retain the parked opt-in solver study; read the project ledger before proposing new work. | 2026-10-03 |
+
+## Active tasks
+
+Open means tracked, not necessarily running. Blockers and decisions still apply.
+
+| Task | Campaign | Status | Priority | Next step |
+|---|---|---|---|---|
+| [Runtime cells&#x27; A100 `single_jit` includes the post-compile warm-up — source-plane 0.642 ms vs a steady 0.267 ms on the same node](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/runtime_cell_single_jit_gpu_warmup.md) | point-source-source | needs-decision | low | Read the task and resolve its explicit human decision before execution. |
+| [Audit timing tests and profiling gates for measurement noise](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/timing_noise_audit.md) | measurement-tools | ready | high | Read the task and select the next bounded step; check current repository claims. |
+| [MGE likelihood_breakdown steps are cumulative and `linear_gaussians` is reported as 0](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/mge_likelihood_breakdown_steps_are_cumulative_an.md) | measurement-tools | ready | low | Read the task and select the next bounded step; check current repository claims. |
+| [A gradient-cost probe: forward vs `value_and_grad` ms/eval and a strict FD check, on any registry cell](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/gradient_cost_probe.md) | measurement-tools | ready | low | Read the task and select the next bounded step; check current repository claims. |
+| [Numba breakdown harness: perturb the instance so the operated-matrix memo cannot hide a step](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/numba_breakdown_harness_memo_blind.md) | measurement-tools | ready | medium | Read the task and select the next bounded step; check current repository claims. |
+| [Search settings-estimation + profiling infrastructure (n_starts / batch_size / n_batch)](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/search_settings_estimation_infrastructure.md) | measurement-tools | ready | normal | Read the task and select the next bounded step; check current repository claims. |
+| [jax_compile/probe.py lost its cell builder with the searches tier — give profiling its own](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/jax_compile_probe_needs_own_cell_builder.md) | measurement-tools | ready | normal | Read the task and select the next bounded step; check current repository claims. |
+| [Linear-solver programme phase 3: GPU/vmap/A100 timing and parity rows for the solver corpus](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/mge_nnls_fix_pyautoarray_571_slam_60.md) | linear-solver | blocked | medium | a release shipping the phase-2 fix (PyAutoArray#595, merged 2026-09-30 pending release; record complete/2026/09/raw-pdip-forward-polish.md) AND the RAL mirror synced via HPCPullPyAuto |
+| [autofit_profiling: bootstrap the repo + general PyAutoFit profiling epic](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/autofit_profiling_bootstrap.md) | autofit | needs-decision | normal | Read the task and resolve its explicit human decision before execution. |
+| [Cluster PointSolver — robustness, analysis settings and performance](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/cluster_pointsolver_speed.md) | cluster-pointsolver | ready | normal | Read the task and select the next bounded step; check current repository claims. |
+| [Profile imaging pixelizations at production over-sampling](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/imaging_production_over_sampling.md) | imaging | ready | normal | Read the task and select the next bounded step; check current repository claims. |
+| [Interferometer decision matrix — fill the last cell (CPU rect 39² at alma_high r5.0) and flip the wiki row to shipped](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_decision_matrix_last_cell.md) | interferometer | needs-decision | normal | Read the task and resolve its explicit human decision before execution. |
+| [Interferometer fixed-mapper searches: reuse the W~ curvature matrix across likelihood calls via `preloads.curvature_matrix`](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_fixed_mapper_curvature_preload.md) | interferometer | needs-slicing | normal | Scope one bounded phase from the full task contract. |
+| [fnnls warm-start memo: stop it slowing scattered evaluation streams (interferometer CPU, autolens_profiling#332)](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_nnls_memo_scattered_stream_guard.md) | interferometer | ready | medium | Read the task and select the next bounded step; check current repository claims. |
+| [Campaign: interferometer streaming (array-free) vs in-memory — memory and time scaling to 2e8 visibilities](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_streaming_scaling.md) | interferometer | ready | high | Read the task and select the next bounded step; check current repository claims. |
+| [Interferometer W~ curvature matrix is FFT-bound on the mask extent: pruned padded FFT and real-space pixel scale on the A100](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/interferometer_w_tilde_fft_size_levers.md) | interferometer | ready | normal | Read the task and select the next bounded step; check current repository claims. |
+| [PointSolver profiling cells: lensed quasar → cluster runtime tier → single/multi-source → multiplane](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/point_solver_profiling_cells.md) | cluster-pointsolver | ready | normal | Read the task and select the next bounded step; check current repository claims. |
+| [Point-source A100 speed-up campaign: profile and optimize with the shared breakdown](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/point_source_image_plane_gpu_breakdown.md) | point-source-image-gpu | needs-decision | high | Read the task and resolve its explicit human decision before execution. |
+| [Point-source source-plane chi-squared speed-up campaign — remaining candidates (phases 1–2e shipped)](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/point_source_source_plane_chi_squared_speed.md) | point-source-source | blocked | low | Await gradient-sampler admission evidence; core campaign already complete. |
+| [Point-source (single-source) CPU campaign — carried leftovers and completion evidence](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/pointsolver_cpu_speed_campaign_remainder.md) | point-source-image-cpu | ready | low | Read the task and select the next bounded step; check current repository claims. |
+| [Assess remaining likelihood bottlenecks after certified solver integration on CPU and GPUs](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/post_certified_solver_likelihood_breakdown.md) | imaging | needs-slicing | normal | Scope one bounded phase from the full task contract. |
+| [profile_lens_aggregator.py cannot run from the autolens_workspace_developer root: no config/ directory](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/profile_lens_aggregator_needs_config_dir.md) | measurement-tools | ready | low | Read the task, reconcile current evidence and select one bounded step. |
+| [jax_profiling/gradient/imaging/pixelization.py: 3.2% of its pin move is unattributed](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/gradient_pixelization_pin_residual_drift.md) | measurement-tools | ready | normal | Read the task, reconcile current evidence and select one bounded step. |
+| [Pair JAX/XLA env vars with measured compile and run times, per backend](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/pair_jax_xla_env_vars_with_measured.md) | measurement-tools | ready | high | Read the task, reconcile current evidence and select one bounded step. |
+| [Quick-update plotting cost — minutes per update, and it is not JAX compile](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/quick_update_plotting_cost.md) | imaging | ready | medium | Read the task, reconcile current evidence and select one bounded step. |
+
+Completed and superseded tasks remain in the [ledger](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/campaigns.yaml).
+
+## Profiling evidence
+
 The cross-project view of the organism's profiling. Each `<lib>_profiling` project repo runs its producers, keeps its results and drift policy, and publishes a `profiling-summary` file. This page reads each file at one resolved commit, validates the exchange contract and shows what the producer says: it recomputes no ratio, combines no timings whose axis or hardware differs, ranks nothing across projects and issues no verdict. Judgment belongs to the Brain's profiling conductor (`/profiling triage`).
 Live page: [https://pyautolabs.github.io/PyAutoPulse/](https://pyautolabs.github.io/PyAutoPulse/).
 
diff --git a/migration.yaml b/migration.yaml
new file mode 100644
index 0000000..b713b4a
--- /dev/null
+++ b/migration.yaml
@@ -0,0 +1,83 @@
+version: 1
+date: '2026-10-03'
+source_repo: PyAutoLabs/PyAutoMind
+source_commit: f21898e042be67afd86cc9c512879b3ad1c9ca9a
+tasks:
+- source: draft/bug/autolens_profiling/runtime_cell_single_jit_gpu_warmup.md
+  destination: tasks/runtime_cell_single_jit_gpu_warmup.md
+  sha256: 8b378f3232fcc138e0c31e182bc7b04678533c9b40b8966c522568213952cf85
+- source: draft/bug/autolens_profiling/timing_noise_audit.md
+  destination: tasks/timing_noise_audit.md
+  sha256: 8f61afc4f7b1bc708121bcc252ab4e05644f09ab337d3db5bc0ecfc7392739d6
+- source: draft/bug/workspaces/mge_likelihood_breakdown_steps_are_cumulative_an.md
+  destination: tasks/mge_likelihood_breakdown_steps_are_cumulative_an.md
+  sha256: bb0ac283af2e89d36a1b3c6092ea1f4539a4a219535f1fa0dd6fc02847b0f3e6
+- source: draft/feature/autolens_profiling/gradient_cost_probe.md
+  destination: tasks/gradient_cost_probe.md
+  sha256: 683afcce80b57fb86b9b265bccdbb6d5076e6204feacc26a453b292059e92f9f
+- source: draft/feature/autolens_profiling/numba_breakdown_harness_memo_blind.md
+  destination: tasks/numba_breakdown_harness_memo_blind.md
+  sha256: 4e756c201e8fcc553be59e739b5116892e3e8df7c890d871d714cda7982ad044
+- source: draft/feature/autolens_profiling/search_settings_estimation_infrastructure.md
+  destination: tasks/search_settings_estimation_infrastructure.md
+  sha256: 71d0d787966c44a8bc0d9b0925e951c942421e0d8150468446aca245e357bd43
+- source: draft/maintenance/autolens_profiling/jax_compile_probe_needs_own_cell_builder.md
+  destination: tasks/jax_compile_probe_needs_own_cell_builder.md
+  sha256: 4e80809d091e489af339882857f3375de60e51a4bdb042d4066ee96f54f60927
+- source: draft/maintenance/autolens_profiling/mass_field_flat_adoption_science_repos.md
+  destination: tasks/mass_field_flat_adoption_science_repos.md
+  sha256: e74045ccd834799ed042d57864ad311dc73ab8fbcb10fd8abb3baead139c52ad
+- source: draft/research/autoarray/mge_nnls_fix_pyautoarray_571_slam_60.md
+  destination: tasks/mge_nnls_fix_pyautoarray_571_slam_60.md
+  sha256: bdf0657cc3f1c76471530f6a9c239859c99653249245232cd2e5f1e52e2c6b8c
+- source: draft/research/autofit/autofit_profiling_bootstrap.md
+  destination: tasks/autofit_profiling_bootstrap.md
+  sha256: 90208326220cbaa16a11411b07df2532e24102e337b0a219e1dfaff8824b4376
+- source: draft/research/autolens_profiling/cluster_pointsolver_speed.md
+  destination: tasks/cluster_pointsolver_speed.md
+  sha256: 996c07763ec0d71b7c0df0de21451c8c3fc0415b8bd55da2503caa92d8bb7534
+- source: draft/research/autolens_profiling/imaging_production_over_sampling.md
+  destination: tasks/imaging_production_over_sampling.md
+  sha256: 67c351e88b7936304c25c5d3f72502600f23fbf30cbad6b1afde95c3c1c3e243
+- source: draft/research/autolens_profiling/interferometer_decision_matrix_last_cell.md
+  destination: tasks/interferometer_decision_matrix_last_cell.md
+  sha256: 16296c06b6e381ba1718654e259811a2c0de07983be30a95aee3e1cbcc0aad4e
+- source: draft/research/autolens_profiling/interferometer_fixed_mapper_curvature_preload.md
+  destination: tasks/interferometer_fixed_mapper_curvature_preload.md
+  sha256: 9c85888f7898332251a8b651dcbb0699c5601c4fd962fa81397a5090e0712288
+- source: draft/research/autolens_profiling/interferometer_nnls_memo_scattered_stream_guard.md
+  destination: tasks/interferometer_nnls_memo_scattered_stream_guard.md
+  sha256: c0045208237e971abc4b08f8ef40fcc37f554cfac4f12ce4ada29aa74f5dfe39
+- source: draft/research/autolens_profiling/interferometer_streaming_scaling.md
+  destination: tasks/interferometer_streaming_scaling.md
+  sha256: 8346fee374c6e7038379306735ae0d1c8311c7bdd2a3eb38f42c49e4a82c4c1a
+- source: draft/research/autolens_profiling/interferometer_w_tilde_fft_size_levers.md
+  destination: tasks/interferometer_w_tilde_fft_size_levers.md
+  sha256: 9700426a4a4af82fd2c6a6af8cacd4a1c77f2a12f3010cd4979a91f29ed57c21
+- source: draft/research/autolens_profiling/point_solver_profiling_cells.md
+  destination: tasks/point_solver_profiling_cells.md
+  sha256: 752cb51d2aecf1f1aca36af59d74b2a1a108216cca7ec4cf1112467f24925135
+- source: draft/research/autolens_profiling/point_source_image_plane_gpu_breakdown.md
+  destination: tasks/point_source_image_plane_gpu_breakdown.md
+  sha256: 2951f69a304eb8909af19dfc7dd4f12223a8e90699faa12d59c32c30a5c490f2
+- source: draft/research/autolens_profiling/point_source_source_plane_chi_squared_speed.md
+  destination: tasks/point_source_source_plane_chi_squared_speed.md
+  sha256: 404d6a175f33c48b6fab0ac6a2dac1bc8651b9748d889a206ecb05b982c2eb37
+- source: draft/research/autolens_profiling/pointsolver_cpu_speed_campaign_remainder.md
+  destination: tasks/pointsolver_cpu_speed_campaign_remainder.md
+  sha256: 396eeaec5c7c73b539beaaa1ba6c926cb5544e11915dba7a1f69199a80b73db5
+- source: draft/research/autolens_profiling/post_certified_solver_likelihood_breakdown.md
+  destination: tasks/post_certified_solver_likelihood_breakdown.md
+  sha256: d9c1ae6441dca7a5da76a9597b71e25e7d38fd82025d1aba6783a1e8318f75a9
+- source: draft/bug/workspaces/profile_lens_aggregator_needs_config_dir.md
+  destination: tasks/profile_lens_aggregator_needs_config_dir.md
+  sha256: 85a3c2faefc51d28dce93731aa0a257e58aedb280607004aaece9566b8604b81
+- source: draft/bug/workspaces/gradient_pixelization_pin_residual_drift.md
+  destination: tasks/gradient_pixelization_pin_residual_drift.md
+  sha256: 8a4f9cbbd939077facb00d87253e778ad99b14f9a789f7bef1be887f52789b3c
+- source: draft/research/autonerves/pair_jax_xla_env_vars_with_measured.md
+  destination: tasks/pair_jax_xla_env_vars_with_measured.md
+  sha256: 66036651646a2af1a172174f2910e8c72a4463a58f19b94bbbd1d4a661433afb
+- source: draft/research/autolens/quick_update_plotting_cost.md
+  destination: tasks/quick_update_plotting_cost.md
+  sha256: 62a01bea5c946bb8a87f0565069b928788ead5dcf1d46e5589722e38308bb13c
diff --git a/pulse/board.py b/pulse/board.py
index e5e25ad..11723e3 100644
--- a/pulse/board.py
+++ b/pulse/board.py
@@ -1,7 +1,7 @@
 """Build the board (``dashboard.md`` + ``dashboard.html``, plus ``badge.json`` and
 the ``state.json`` organ-cockpit feed) from the ingested snapshots.
 
-The first section is one row per registered project: scope, evidence time,
+The campaign control room comes first. The evidence section has one row per registered project: scope, evidence time,
 last fetch (with the resolved commit), coverage, and three separate columns
 for the three notions the spec keeps apart — **integrity** (transport and
 schema: ok / invalid / unavailable / unsupported / cached), **freshness**
@@ -31,7 +31,7 @@
 from datetime import UTC, datetime
 from pathlib import Path
 
-from pulse import ORGAN_ROOT
+from pulse import ORGAN_ROOT, campaigns
 from pulse import summary as summary_mod
 from pulse.ingest import Snapshot
 
@@ -337,7 +337,7 @@ def _md_detail(s: Snapshot, now: str | None = None) -> list[str]:
     return out
 
 
-def render_markdown(views, now: str | None = None) -> str:
+def render_markdown(views, now: str | None = None, campaign_data: dict | None = None) -> str:
     now = now or _utc_now()
     out = [
         "# PyAutoPulse — profiling dashboard",
@@ -345,6 +345,7 @@ def render_markdown(views, now: str | None = None) -> str:
         "<!-- Generated by `bin/pyauto-pulse board` from registry.yaml, receipts/ and "
         "snapshots/. Do not edit by hand. -->",
         "",
+        campaigns.markdown(campaign_data if campaign_data is not None else campaigns.load()),
         LEDE,
         f"Live page: [{PAGES_URL}]({PAGES_URL}).",
         "",
@@ -410,10 +411,23 @@ def render_markdown(views, now: str | None = None) -> str:
 background:var(--bg);color:var(--fg);border:1px solid var(--line);border-radius:6px;
 padding:3px 6px;cursor:pointer;word-break:break-all}
 button.copied{border-color:var(--ok)}
+.checkin textarea{display:block;width:100%;font:inherit;font-size:14px;line-height:1.5;background:var(--card);color:var(--fg);border:1px solid var(--line);border-radius:8px;padding:12px;margin:8px 0;resize:vertical}
+#copy-checkin{background:var(--accent);color:var(--bg);border:0;border-radius:6px;padding:10px 16px;cursor:pointer;font-weight:600}
+#copy-status{margin-left:12px}
 code{color:var(--accent);word-break:break-all}
 """
 
 JS = """
+var copyCheckin=document.getElementById('copy-checkin');
+if(copyCheckin){copyCheckin.addEventListener('click',async function(){
+  var field=document.getElementById('checkin-prompt'), status=document.getElementById('copy-status');
+  try {
+    if(navigator.clipboard && window.isSecureContext){await navigator.clipboard.writeText(field.value)}
+    else {field.focus();field.select();if(!document.execCommand('copy')){throw new Error('copy unavailable')}}
+    status.textContent='Copied';
+  } catch(error){field.focus();field.select();status.textContent='Select and copy the prompt above.'}
+});}
+
 document.querySelectorAll('button[data-copy]').forEach(function(b){
   b.addEventListener('click',function(){
     var t=b.getAttribute('data-copy');
@@ -537,7 +551,7 @@ def _html_detail(s: Snapshot, now: str | None = None) -> str:
     return "".join(parts)
 
 
-def render_html(views, now: str | None = None) -> str:
+def render_html(views, now: str | None = None, campaign_data: dict | None = None) -> str:
     now = now or _utc_now()
     rows = []
     for s in views:
@@ -566,7 +580,8 @@ def render_html(views, now: str | None = None) -> str:
         f"<style>{CSS}</style></head><body>"
         "<header class='hero'><h1>PyAuto<span>Pulse</span></h1>"
         "<p>Measure. Trace. Compare.</p></header>"
-        f"<main><p class='lede'>{lede}</p>"
+        f"<main>{campaigns.render_html(campaign_data if campaign_data is not None else campaigns.load())}"
+        f"<p class='lede'>{lede}</p>"
         f"<div class='stats'>{tiles}</div>"
         "<h2 id='projects'>Projects</h2>"
         "<div class='tablewrap'><table><thead><tr><th>Project</th><th>Scope</th>"
diff --git a/pulse/campaigns.py b/pulse/campaigns.py
new file mode 100644
index 0000000..3170056
--- /dev/null
+++ b/pulse/campaigns.py
@@ -0,0 +1,191 @@
+"""Human-maintained profiling intent; never infer scientific outcomes from timings."""
+
+from __future__ import annotations
+
+import hashlib
+import html
+import json
+import re
+from pathlib import Path
+from urllib.parse import urlparse
+
+import yaml
+
+from pulse import ORGAN_ROOT
+
+URL = "https://github.com/PyAutoLabs/PyAutoPulse/blob/main/"
+STATUSES = {
+    "active",
+    "ready",
+    "running",
+    "blocked",
+    "parked",
+    "needs-decision",
+    "needs-slicing",
+    "complete",
+    "superseded",
+}
+CLOSED = {"complete", "superseded"}
+PROMPT = """Use PyAutoPulse as the home for all profiling work in this chat. Read PyAutoPulse/AGENTS.md and CHECKIN.md, then campaigns.yaml, the linked tasks and current project campaign ledgers. Check every open campaign and task for updates since the last check-in: results, jobs, PRs, releases, blockers and the next bounded step. Distinguish verified updates from stale or unavailable evidence; do not infer success from missing data. Give me one concise campaign-by-campaign summary and a proposed priority order. Apply any campaign direction or ideas I supply before or after this prompt; otherwise cover everything. Keep the overall check-in even when I focus one campaign. Update the Pulse ledger with verified facts and dated source links, regenerate the board, and use the repository workflow to persist changes. Use Brain's profiling conductor for scientific judgement and project drivers for execution. Do not launch new compute, change defaults or baselines, or bypass a human/release gate merely to check in. Continue managing subsequent profiling requests in this chat. Optional direction: [leave blank, name a campaign, or suggest an idea]."""
+
+
+class CampaignError(ValueError):
+    """The control-room ledger is invalid or incomplete."""
+
+
+def load(root: Path = ORGAN_ROOT) -> dict:
+    root = Path(root)
+    path = root / "campaigns.yaml"
+    if not path.is_file():
+        raise CampaignError(f"{path}: missing campaign ledger")
+    try:
+        data = yaml.safe_load(path.read_text())
+    except yaml.YAMLError as exc:
+        raise CampaignError(str(exc)) from exc
+    if not isinstance(data, dict) or data.get("version") != 1:
+        raise CampaignError("campaign ledger must have version 1")
+    for kind in ("campaigns", "tasks"):
+        rows = data.get(kind)
+        if not isinstance(rows, list):
+            raise CampaignError(f"{kind} must be a list")
+        seen = set()
+        for row in rows:
+            if not isinstance(row, dict):
+                raise CampaignError(f"{kind}: expected a mapping")
+            for field in ("id", "title", "status", "updated", "next"):
+                if not isinstance(row.get(field), str) or not row[field].strip():
+                    raise CampaignError(f"{kind}: missing text {field}")
+            if not re.fullmatch(r"[a-z0-9][a-z0-9_-]*", row["id"]) or row["id"] in seen:
+                raise CampaignError(f"{kind}: invalid/duplicate id {row['id']}")
+            seen.add(row["id"])
+            if row["status"] not in STATUSES:
+                raise CampaignError(f"{row['id']}: unknown status")
+            for field in ("evidence", "issue"):
+                if field in row and (
+                    not isinstance(row[field], str)
+                    or urlparse(row[field]).scheme != "https"
+                    or not urlparse(row[field]).netloc
+                ):
+                    raise CampaignError(f"{row['id']}: {field} must be an HTTPS URL")
+    ids = {c["id"] for c in data["campaigns"]}
+    for task in data["tasks"]:
+        if task.get("campaign") not in ids:
+            raise CampaignError(f"{task['id']}: unknown campaign")
+        relative = task.get("path", "")
+        if not isinstance(relative, str):
+            raise CampaignError(f"{task['id']}: task path must be text")
+        path = root / relative
+        if (
+            not isinstance(relative, str)
+            or not relative.startswith("tasks/")
+            or not path.resolve().is_relative_to((root / "tasks").resolve())
+            or not path.is_file()
+        ):
+            raise CampaignError(f"{task['id']}: missing or unsafe task path")
+    return data
+
+
+def marker(data: dict) -> str:
+    digest = hashlib.sha256(json.dumps(data, sort_keys=True).encode()).hexdigest()
+    return f"<!-- pulse-campaigns:{digest} -->"
+
+
+def _md(value) -> str:
+    return html.escape(str(value)).replace("|", "\\|").replace("\n", " ")
+
+
+def markdown(data: dict) -> str:
+    rows = [
+        marker(data),
+        "## Check in on all profiling work",
+        "",
+        "Copy this into one chat. Add a campaign focus or idea before or after it, or leave it unchanged.",
+        "",
+        "```text",
+        PROMPT,
+        "```",
+        "",
+        f"Last check-in: {data.get('last_checkin') or 'not recorded yet'}. Ledger dates are review dates, not measurement freshness.",
+        "",
+        "## Active campaigns",
+        "",
+        "| Campaign | Status | Open tasks | Next step | Reviewed |",
+        "|---|---|---:|---|---|",
+    ]
+    for c in data["campaigns"]:
+        if c["status"] in CLOSED:
+            continue
+        n = sum(t["campaign"] == c["id"] and t["status"] not in CLOSED for t in data["tasks"])
+        title = f"[{_md(c['title'])}]({c['evidence']})" if c.get("evidence") else _md(c["title"])
+        rows.append(f"| {title} | {c['status']} | {n} | {_md(c['next'])} | {c['updated']} |")
+    rows += [
+        "",
+        "## Active tasks",
+        "",
+        "Open means tracked, not necessarily running. Blockers and decisions still apply.",
+        "",
+        "| Task | Campaign | Status | Priority | Next step |",
+        "|---|---|---|---|---|",
+    ]
+    for t in data["tasks"]:
+        if t["status"] not in CLOSED:
+            rows.append(
+                f"| [{_md(t['title'])}]({URL}{t['path']}) | {t['campaign']} | {t['status']} | {_md(t.get('priority', 'normal'))} | {_md(t['next'])} |"
+            )
+    rows += [
+        "",
+        "Completed and superseded tasks remain in the [ledger](" + URL + "campaigns.yaml).",
+        "",
+        "## Profiling evidence",
+        "",
+    ]
+    return "\n".join(rows)
+
+
+def render_html(data: dict) -> str:
+    def e(value):
+        return html.escape(str(value), quote=True)
+
+    campaigns = []
+    for c in data["campaigns"]:
+        if c["status"] in CLOSED:
+            continue
+        n = sum(t["campaign"] == c["id"] and t["status"] not in CLOSED for t in data["tasks"])
+        title = (
+            f'<a href="{e(c["evidence"])}">{e(c["title"])}</a>'
+            if c.get("evidence")
+            else e(c["title"])
+        )
+        campaigns.append(
+            f"<tr><td>{title}</td><td>{e(c['status'])}</td><td>{n}</td><td>{e(c['next'])}</td><td>{e(c['updated'])}</td></tr>"
+        )
+    tasks = []
+    for t in data["tasks"]:
+        if t["status"] in CLOSED:
+            continue
+        issue = f' · <a href="{e(t["issue"])}">issue</a>' if t.get("issue") else ""
+        tasks.append(
+            f'<tr><td><a href="{URL}{e(t["path"])}">{e(t["title"])}</a>{issue}</td><td>{e(t["campaign"])}</td><td>{e(t["status"])}</td><td>{e(t.get("priority", "normal"))}</td><td>{e(t["next"])}</td></tr>'
+        )
+
+    def table(headers, rows):
+        return (
+            '<div class="tablewrap"><table><thead><tr>'
+            + "".join(f"<th>{h}</th>" for h in headers)
+            + "</tr></thead><tbody>"
+            + "".join(rows)
+            + "</tbody></table></div>"
+        )
+
+    return (
+        marker(data) + '<section class="checkin"><h2>Check in on all profiling work</h2>'
+        "<p>One prompt, one ongoing chat. Add a campaign focus or idea here, or tell the chat before or afterwards.</p>"
+        f'<label for="checkin-prompt">Your check-in prompt</label><textarea id="checkin-prompt" rows="8">{e(PROMPT)}</textarea>'
+        '<button type="button" id="copy-checkin">Copy check-in prompt</button><span id="copy-status" role="status" aria-live="polite"></span>'
+        f'<p class="muted">Last check-in: {e(data.get("last_checkin") or "not recorded yet")}. Ledger dates are review dates, not measurement freshness.</p></section>'
+        '<h2 id="campaigns">Active campaigns</h2>'
+        + table(["Campaign", "Status", "Open tasks", "Next step", "Reviewed"], campaigns)
+        + '<h2 id="tasks">Active tasks</h2><p>Open means tracked, not necessarily running. Blockers and decisions still apply.</p>'
+        + table(["Task", "Campaign", "Status", "Priority", "Next step"], tasks)
+        + f'<p>Completed and superseded tasks remain in the <a href="{URL}campaigns.yaml">ledger</a>.</p><h2 id="evidence">Profiling evidence</h2>'
+    )
diff --git a/pulse/cli.py b/pulse/cli.py
index 6ba6738..e107c35 100644
--- a/pulse/cli.py
+++ b/pulse/cli.py
@@ -9,7 +9,7 @@
 import sys
 from pathlib import Path
 
-from pulse import ORGAN_ROOT, board, ingest, registry
+from pulse import ORGAN_ROOT, board, campaigns, ingest, registry
 from pulse import summary as summary_mod
 
 
@@ -157,6 +157,14 @@ def _check_state(path: Path) -> list[str]:
 
 def cmd_check(args) -> int:
     problems = []
+    try:
+        campaign_data = campaigns.load()
+    except campaigns.CampaignError as exc:
+        print(f"FAIL campaigns: {exc}")
+        return 1
+    print(
+        f"ok   campaigns: {len(campaign_data['campaigns'])} campaigns, {len(campaign_data['tasks'])} tasks"
+    )
     try:
         body_map = registry.load_body_map(args.mind)
         print(f"ok   body map: {registry.body_map_path(args.mind)}")
@@ -202,6 +210,10 @@ def cmd_check(args) -> int:
         problems.append("dashboard.md / dashboard.html missing")
         print("FAIL dashboard: dashboard.md / dashboard.html missing — run `pyauto-pulse board`")
     else:
+        expected = campaigns.marker(campaign_data)
+        if expected not in md.read_text() or expected not in page.read_text():
+            problems.append("campaign dashboard stale — run pyauto-pulse board")
+            print("FAIL campaigns: dashboard does not match the campaign ledger")
         recorded = board.markers(md.read_text())
         names = {i.instance for i in instances}
         stale = sorted(
diff --git a/tasks/autofit_profiling_bootstrap.md b/tasks/autofit_profiling_bootstrap.md
new file mode 100644
index 0000000..1584d95
--- /dev/null
+++ b/tasks/autofit_profiling_bootstrap.md
@@ -0,0 +1,73 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autofit/autofit_profiling_bootstrap.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# autofit_profiling: bootstrap the repo + general PyAutoFit profiling epic
+
+Type: research
+Target: PyAutoFit
+Repos:
+- PyAutoFit
+- autofit_profiling
+Themes:
+- profiling
+- graphical-ep
+Difficulty: large
+Autonomy: supervised
+Priority: normal
+Status: formalised
+Consequence: glance
+Witness: The `autofit_profiling` repo exists (created only after the human's interactive confirmation) with the ported `ep/`, `graphical/` and analytic-benchmark packages reproducing their committed baseline numbers, and epic 1's issue is open with a measured, ranked bottleneck table for a single `search.fit` on a fast likelihood; epic 2 is not opened.
+Review-minutes: 3
+Unattended: ready
+Campaign: research/graphical_ep/ep_campaign.md (Phase 6)
+Filed: 2026-08-19 (backfilled from git)
+
+## Why
+
+For the IC50 and cosmology use cases the individual Factor fits will end up
+fast; the bottleneck becomes PyAutoFit itself — wrapping up a fit on one
+factor, writing its output, starting the next, building the cavity
+distributions. The 2026-07 scoping baselines already show this shape: EP is
+~11× slower than graphical at every N, with per-factor visualisation alone at
+~48% of EP runtime and `scipy.stats.truncnorm` prior transforms at 25–38%
+(`research/graphical_ep/ep_scoping.md`, `graphical_scoping.md`). Slowdowns in
+`non_linear/` (priors, paths, output) accumulate per factor fit and *become*
+the EP bottleneck — so PyAutoFit needs a general profiling assessment, not
+just an EP one.
+
+## Deliverable 1 — the `autofit_profiling` repo
+
+A new GitHub repo mirroring `autolens_profiling`'s responsibility, but for
+all things PyAutoFit. **Repo creation is human-gated: confirm interactively
+with James before creating anything on GitHub.** Seed it by porting the
+existing baseline packages from `autofit_workspace_developer/` (`ep/`,
+`graphical/`, and the analytic benchmark, shipped 2026-09-02 as
+`complete/2026/09/analytic-gaussian-benchmark.md`) so results,
+harnesses and hotspot files live in one place with committed history, the way
+autolens_profiling does for likelihoods.
+
+## Deliverable 2 — two epics, sequenced (created at start_dev time, one at a
+time — never bulk-issued)
+
+1. **General PyAutoFit profiling** (first): a full profiling assessment of a
+   single `search.fit` — prior transforms, `non_linear` wrapper overhead
+   (paths setup, samples-to-CSV, visualisation, latent samples,
+   search-internal cleanup), output I/O — on a fast likelihood so the
+   overhead has nowhere to hide. Ranked findings filed as one-PR prompts.
+2. **EP-loop profiling** (second, follows on): the EP-specific overhead —
+   cavity construction, message updates, `factor_graph.optimise`
+   orchestration — building directly on the ranked sub-task list already in
+   `ep_scoping.md` (viz suppression, truncnorm fast path, folder-cleanup
+   skip, orchestration cProfile, parallel local fits, sampler reuse). Most
+   per-fit wins from epic 1 multiply by M·N under EP, which is why epic 1
+   lands first.
+
+## Acceptance
+
+- Repo exists with the ported baselines reproducing their committed numbers.
+- Epic 1 issue open with a measured, ranked bottleneck table for a single
+  search.fit; epic 2 issue opened only after epic 1's top findings ship.
diff --git a/tasks/cluster_pointsolver_speed.md b/tasks/cluster_pointsolver_speed.md
new file mode 100644
index 0000000..aab0238
--- /dev/null
+++ b/tasks/cluster_pointsolver_speed.md
@@ -0,0 +1,190 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autolens_profiling/cluster_pointsolver_speed.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Cluster PointSolver — robustness, analysis settings and performance
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+- autolens_workspace_test
+- PyAutoLens
+- PyAutoArray
+- PyAutoGalaxy
+Themes:
+- cluster
+- point-source
+- profiling
+- jax
+Difficulty: large
+Autonomy: supervised
+Priority: normal
+Status: formalised
+Consequence: judge
+Review-minutes: 20
+Unattended: ready
+Epic: cluster-pointsolver-speed
+Filed: 2026-09-26
+Parent-record: complete/2026/09/point-source-cpu-p3.md
+
+## Status and ownership — approved 2026-10-01
+
+Canonical planning ledger for epic `cluster-pointsolver-speed` (identifier
+retained; display title broadened). `autolens_profiling` is the execution/evidence
+home for both “make autolens fast” and “which settings make the analysis robust”.
+This expands the unstarted cluster speed campaign; it does not restart completed
+single-source work. The profiling repo's `wiki/campaigns/cluster_pointsolver.md`
+is the earlier speed-only survey, to be reconciled when the first campaign task
+starts; this Mind contract governs the newly approved scope in the meantime.
+
+Transferred from Source & Cluster arc:
+- `draft/bug/autolens/point_solver_error_bisect_health.md`: remaining phase-1
+  overflow, accuracy/identity and regression obligations, historical-bisect
+  limitation disposition. Completed 1a–1d records remain in their original arc.
+- `draft/research/autolens_profiling/point_solver_profiling_cells.md`: old phase-2
+  coverage gaps, reconciled with current profiling taxonomy and the already
+  completed single-source campaign before any issue is filed.
+- Forward ShapeSolver rehabilitation/retirement and its finite-source convergence
+  tests from arc phase 6. Primary pixel-based area magnification stays in the arc.
+
+No task in this programme is issued by this transfer. Issue ONE bounded task at
+a time as its predecessor nears shipping; not the umbrella and not a bulk queue.
+
+## Working contract
+
+1. **Robustness/settings evidence belongs in autolens_profiling.** Measure initial
+   grid scale, extent, target precision, capacity, backend/dtype and applicable
+   source/redshift regimes against independent numerical references. Publish the
+   supported settings and their tested domain, unresolved outcomes, cost and
+   provenance together. Shared library repairs remain general, not cluster-only.
+2. **Numerical integration guarantees belong in autolens_workspace_test.** Distil
+   every accepted repair/settings claim into bounded reproducible regressions:
+   positions, per-image coverage, overflow, unresolved multiplicity, eager/JIT/vmap
+   and registered-tracer/gradient behavior where affected. Name and wire a CI lane:
+   cheap deterministic witnesses in required PR smoke; expensive coverage in the
+   scheduled/release suite. A standalone opt-in diagnostic with no CI consumer is
+   not completion of this regression obligation. Existing 1c/1d cells remain
+   historical evidence in workspace_test; do not move or duplicate them wholesale.
+3. **Performance measurement can begin before all repairs finish.** Label failing
+   cases diagnostic/invalid and retain their timings, but exclude them from
+   accepted speed-up and robust-settings claims. Matching the old likelihood is
+   insufficient: require independent accuracy/coverage, no silent truncation and
+   relevant derivative/batch checks before accepting an optimisation. Record
+   compile, first-call and warmed costs separately on pinned hardware.
+4. **Finite acceptance target.** Specify supported workloads/tolerances first;
+   detect overflow, establish per-image accuracy/coverage, expose unresolved cases,
+   and verify backend behavior. Do not use “trusted everywhere” as an unbounded
+   exit condition or silently change scientific tolerances to win speed.
+5. **Handoffs.** Profiling evidence → bounded library repair → CI regression →
+   settings recommendation. Downstream Source & Cluster tasks depend on the
+   concrete capability they use, not completion of this whole programme.
+
+## Next task selection and workstreams
+
+Next production candidate: observable containment-overflow contract under eager,
+JIT and vmap, including caller invalid-result behavior and CI witnesses. Recheck
+claims and approve the API design through start_dev before issuing it. The earlier
+streaming #600 Array claim has closed; that is not a standing conflict waiver.
+Representative cluster data and baseline profiling below can be scoped without
+waiting for every accuracy question; failing configurations remain labelled.
+
+Accuracy/identity work uses 1c/1d's counterexamples: three non-converged false
+accepts from the local correction screen; adding convergence loses closest-cusp
+coverage. Investigate safeguarded refinement/nonlinear error control, not another
+unqualified distance/residual rule. Forward ShapeSolver work requires its own
+bounded current-code audit, backend dispatch, per-image area/convergence evidence
+and CI tests, or an explicit retirement decision. Preserve the old phase-6
+requirement to compare any promoted forward implementation with independent
+simulated area truth/pixel-based results; it no longer blocks the primary API.
+
+Completed evidence: `complete/2026/09/point-solver-error-audit.md`,
+`complete/2026/09/point-solver-padding-backend.md`,
+`complete/2026/10/point-solver-duplicate-policy.md`,
+`complete/2026/10/point-solver-image-accuracy.md`.
+
+## User-approved expansion (verbatim)
+
+Ok  I agreem autolens_profiling is evolving into "make autolens fast" but also "these are the settings I need to ensure the analysis is robust" so this work is also more and more belonging there, albeit we need numerical integration tests in autolens_workspace_test which ensure this all makes it way into CI. Update accordingly and then wrap up
+
+## User request (verbatim)
+
+Lets push through all work on single source, working in autolens_profiling/scritps/point_source, we will do cluster use case as as eparate epic so maybe move this as part of an intake that will do that whole thing starting from working out the data and likelihood_breakdown.
+
+## Profiling workstream — cluster data and a released-code baseline
+
+No lever is ranked until this exists. Deliverables:
+
+1. **Work out the cluster data.** Pick one or more representative cluster datasets and justify them
+   against real cluster modelling in `@autolens_workspace` (`scripts/cluster/`). For each, record:
+   - the number of sources and source planes / redshifts;
+   - the number of multiple images per source;
+   - the lens model components (BCG, cluster-scale halo(s), member galaxies — profile types and
+     counts, e.g. dPIE / NFW);
+   - the solver grid (extent, `pixel_scale_precision`) production cluster fits actually use.
+
+   The current profiling case (`dataset/cluster/simple`, gitignored and auto-simulated: 13 mass
+   components, 3 planes, 2 sources) is a starting point, not a verdict.
+2. **Audit and refresh `@autolens_profiling` `scripts/cluster/likelihood_breakdown/`.**
+   `image_plane.py` and `source_plane.py` both exist; check them against the point-source
+   instrument (`scripts/point_source/likelihood_breakdown/` plus the shared
+   `scripts/misc/likelihood_breakdown/{timing,provenance,_profile_cli}.py`). Then bring them to the
+   campaign measurement contract:
+   - fused production likelihood control;
+   - separated compile vs first vs warmed timings;
+   - interleaved medians + CI;
+   - `source_revisions`, thread environment, XLA memory and FLOPs recorded;
+   - a per-step / per-component split (step 0 vs refinement, deflections vs bookkeeping, per
+     source).
+   - **Precision caveat:** the current cell runs `pixel_scale_precision=0.01` (to keep compile at
+     minutes), not production `0.001`. Measure both, or justify the choice explicitly.
+3. **Baseline on released code** (2026.9.26.1 or later, which carries point-source p2 + p3) on a
+   pinned RAL CPU node (Xeon 8490H, `--nodelist`; check `sinfo` first), with an A100 row. Update
+   the README dashboards. Record the per-step wall-time split, which then ranks phase 2+.
+
+## Carried evidence (from point-source p1–p3; records `complete/2026/09/point-source-cpu-p{1,2,3}.md`, ledger `autolens_profiling/results/notes/point_source_cpu_campaign.md`)
+
+- **Two-source cluster solved likelihood, per phase:**
+  - p1 RAL baseline (Xeon 8490H): fused plain 128.05 ms, fused solved 134.52 ms, per-source solves
+    110 / 115 ms. Per-source solves exceed the fused call, a fusion-boundary effect.
+  - p2 vertex-dedup removal (RAL job 350582, EPYC 7763): 155.22 → 78.32 ms (1.98×). The control is
+    +15–20 % from the host alone, so pin the node.
+  - p3 static step-0 lattice (RAL job 350636, Xeon 8490H pinned): 47.36 → 9.085 ms (5.21×); 3.9–4.5×
+    with constant folding on.
+  - FLOPs: 184.2M → 139.4M → 39.5M.
+  - XLA temp memory: 91.6 → 22.6 MB (A100 42.1 → 6.35 MB).
+- **Lattice:** the 200×200 @ 0.7″ cluster step-0 lattice now deflects **46 516 of 276 507** unique
+  vertices.
+- **Where the cost is now (FLOP estimate, not measured):** deflections of the 13-component lens
+  dominate every cluster step. Step 0 is ≈ 51 % of cluster FLOPs.
+- **Lever: dPIE/NFW deflection cost.** This is likely a PyAutoGalaxy phase, separately scoped.
+  `@autolens_profiling` `scripts/lens/deflections/` is **NumPy-only** and has **no dPIE spec**, so a
+  JAX deflection cell with dPIE/NFW specs is a prerequisite for measuring it.
+- **Lever: grid-extent guidance.** Stronger at cluster scale, where it acts on the ≈ 51 % step-0
+  share plus containment. It needs image-completeness evidence across a prior, with each extent
+  reported as its own configuration and never substituted into a speed row.
+- **Unmeasured control:** direct deflections on the full 276 507-point cluster input. The old
+  "cluster solve is at its deflection bound" claim was extrapolated, not measured.
+- **Single-source phase 4a** (solver-config sweep: initial scale × precision,
+  `MAX_CONTAINING_SIZE`, extent) may produce methods or findings reusable here. Read its ledger
+  section before planning phase 2.
+
+## Measurement contract
+
+This inherits the point-source campaign contract (phase-4 prompt, "Campaign contract (2026-09-19)"):
+- exact revisions recorded;
+- interleaved A/B on identical hardware with distinct function objects + `jax.clear_caches()`;
+- a fused production control;
+- a correctness gate before any speed claim: bit-identical likelihood, solved positions,
+  multi-source / multiplane coverage, JVP / vmap parity;
+- negative results recorded;
+- one bounded phase per issue and PR.
+
+## Later performance candidates (re-rank after the baseline)
+
+2. Deflection-cost lever (JAX dPIE/NFW deflection cell first; PyAutoGalaxy change separately scoped).
+3. Cluster grid-extent guidance with completeness evidence.
+4. Whatever the phase-1 split exposes (per-source batching, multiplane bookkeeping, capacity).
diff --git a/tasks/gradient_cost_probe.md b/tasks/gradient_cost_probe.md
new file mode 100644
index 0000000..4621eb8
--- /dev/null
+++ b/tasks/gradient_cost_probe.md
@@ -0,0 +1,81 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/feature/autolens_profiling/gradient_cost_probe.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# A gradient-cost probe: forward vs `value_and_grad` ms/eval and a strict FD check, on any registry cell
+
+Type: feature
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+Themes:
+- jax-gradient
+- profiling
+Difficulty: small
+Autonomy: safe
+Priority: low
+Status: draft
+Consequence: glance
+Witness: the probe runs on CPU and prints a forward/grad ms-per-eval table and an FD table for a named cell
+Review-minutes: 3
+Unattended: ready
+Filed: 2026-09-04
+
+Build a standing profiling tool that answers, for any profiling cell, **what a gradient
+evaluation costs and whether it is correct**. Split out of
+`draft/feature/autolens_profiling/gradient_slam_mass_pix_target.md` on 2026-09-04 by human
+decision: the `gradient-slam-baseline` epic asks whether gradients reduce the *number* of
+inference steps, not what a gradient costs, so the probe is not a science phase and gates
+nothing. It is a tool — keep it in mind, and run it whenever someone wants the number.
+
+## Why
+
+The only rectangular kernel-CDF gradient cost datum that exists anywhere is a **CPU** one:
+the ~17× `value_and_grad`-over-forward anomaly recorded in
+autolens_workspace_developer#117 and `searches_minimal/pix_prodigy_findings.md`. It has
+**never been re-measured on an A100**, so nobody knows whether it is a real cost of the
+mesh likelihood or an artifact of the CPU path. That is a question worth being able to
+answer on demand, on any cell, rather than a question worth a science phase — hence a
+script and a submit script, not a programme.
+
+## What to build
+
+### 1. The probe — `scripts/misc/searches/probe_gradient_cost.py`
+
+A CLI tool taking `--cell <model_type>` — **default `mass_pix`**, and accepting **any cell
+registered in `scripts/misc/searches/_targets.py`**, so it is not tied to one target. For
+the named cell it reports:
+
+- the **forward** likelihood cost in ms/eval;
+- the **`value_and_grad`** cost in ms/eval;
+- **their ratio** — the number the 17× anomaly is about;
+- the **jit compile time** for each of the two;
+- a **strict finite-difference check on every free parameter of the cell** — each
+  FD/analytic pair agreeing to a declared tolerance, no parameter skipped, no NaN or
+  non-finite entry in either gradient, reported as a per-parameter table.
+
+Follow the FD-certification pattern of
+`@autolens_workspace_test/scripts/imaging/jax_grad/pixelization.py` — that script is the
+reference for how the check is set up and what it asserts.
+
+The number of free parameters is the cell's own, read from the model rather than hardcoded,
+so the FD table is complete for whatever cell is asked for.
+
+### 2. The submit — `@autolens_profiling/hpc/batch_gpu/`
+
+One submit script for **one A100 task** running the probe on the requested cell, with the
+usual `WALL-BASIS` block. This is what turns "the 17× has never been measured on a GPU"
+into a one-command answer.
+
+## Out of scope
+
+Anything in the `gradient-slam-baseline` ledger
+(`@autolens_profiling/results/notes/gradient_slam/LEDGER.md`) — the `mass_pix` target
+itself, its Nautilus and Prodigy drivers, their submits, and the Cortex phases 21–23 that
+run them. This prompt adds a tool beside them and changes none of them. No library code in
+PyAutoArray / PyAutoGalaxy / PyAutoLens: if the probe finds a real gradient-cost or
+gradient-correctness fault, that is a bug filed in the Mind, not fixed here. No RAL
+submission as part of the task — the submit script is written, not run.
diff --git a/tasks/gradient_pixelization_pin_residual_drift.md b/tasks/gradient_pixelization_pin_residual_drift.md
new file mode 100644
index 0000000..24fd1ab
--- /dev/null
+++ b/tasks/gradient_pixelization_pin_residual_drift.md
@@ -0,0 +1,90 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/bug/workspaces/gradient_pixelization_pin_residual_drift.md
+Migrated-at: 2026-10-03
+
+Original task retained verbatim below. Current scheduling state: `campaigns.yaml`. Historical paths: `migration.yaml`.
+
+---
+
+# jax_profiling/gradient/imaging/pixelization.py: 3.2% of its pin move is unattributed
+
+Type: bug
+Target: workspaces
+Repos:
+- @autolens_workspace_developer
+- @PyAutoArray
+Themes:
+- pixelization
+- profiling
+- jax-gradient
+Difficulty: medium
+Autonomy: supervised
+Priority: normal
+Status: draft
+Consequence: glance
+Witness: The bisect names one commit whose parent and self, both built and measured (never attributed from commit messages), differ by the residual 3.2% in the eager `figure_of_merit` of `jax_profiling/gradient/imaging/pixelization.py`; that attribution is written into the file's provenance comment beside the #490 rows — or, if the move is a regression, the fix restores the fiducial to a value reproducing its pin to the tolerance the sibling jit cell already meets.
+Review-minutes: 3
+Unattended: ready
+Filed: 2026-08-26
+
+## What happened
+
+Regenerating the adaptive-rectangular pins after PyAutoArray#490 (mirrored
+bilinear row weights + round-off-dependent cell assignment in the shared
+adaptive rectangular mapper) turned up one fiducial whose pin does **not**
+reproduce even before that fix.
+
+`@autolens_workspace_developer/jax_profiling/gradient/imaging/pixelization.py`
+(HST, 28x28 `RectangularBilinearAdaptDensity`, Sersic + Isothermal lens,
+`Constant` regularization) asserted `EXPECTED_LOG_EVIDENCE_HST =
+-66270.78281169113`, pinned 2026-04-24 (`cfa5378`). Measured 2026-08-26 on
+library `main`s, both legs on one host differing only in the #490 commit:
+
+```
+72fb01d1^ (pre-fix)   -84633.30874141103    18362  (27.7%) below the pin
+72fb01d1  (fixed)     -64156.79838547805     2114   (3.2%) above the pin
+```
+
+#490 accounts for +20476 of that. The residual **3.2% is older drift and is
+not attributed**. The pin was re-pointed to the measured post-#490 value in the
+same change (with the provenance recorded in the file); this prompt owns
+explaining the 3.2%.
+
+## Why it is worth a look rather than a shrug
+
+- The input side is ruled out. `jax_profiling/dataset/imaging/hst` is a pure
+  rename at `f8a5cef` (`R100`, identical blobs), and the script's model, mask,
+  over-sampling and mesh shape are byte-identical to the pinning commit — only
+  the path strings changed. The move is entirely library-side.
+- It is **not** the mesh-name trap. `RectangularRTUAdaptDensity` at the same
+  fiducial gives `-86488.3491981348` pre-fix, further from the pin, not closer,
+  so the 2026-08-21 relabel to `Bilinear` (`08d5d86`) was right.
+- The sibling `jax_profiling/jit/imaging/pixelization.py` (MGE lens, 35x35, same
+  mesh family, same dataset directory) still reproduces its own 2026-05-11 pin
+  **bit-for-bit** pre-#490 — measured in the same session. So whatever moved the
+  gradient fiducial left the shared adaptive-mesh path alone, and points at
+  something specific to Sersic + Isothermal + `Constant` reg, or at a change
+  between 2026-04-24 and 2026-05-11.
+
+## Scope
+
+1. Bisect the eager `figure_of_merit` for this fiducial from 2026-04-24 to now.
+   Build the old checkouts and measure — do not attribute from commit messages
+   (the trap recorded in `complete/2026/08/pixelization-eager-jit-divergence.md`).
+   A `git worktree add <scratch> <sha>^ --detach` plus a PYTHONPATH substitution
+   of the one repo is minutes per step, and decisive.
+2. Decide what the answer means: a deliberate improvement nobody re-pinned, or a
+   silent regression in the Sersic / Isothermal / `Constant` path that a stale
+   pin has been hiding since April.
+3. If it is a regression, that is the bug — fix it and re-measure the pin again.
+   If it is deliberate, record the attribution in the file's provenance comment
+   next to the #490 rows already there.
+
+## Notes
+
+- Related but distinct: `draft/bug/workspaces/jax_likelihood_pins_stale_by_1e4.md`
+  covers `autolens_workspace_test`'s marginal (~1.2e-4) pin drift. This one is
+  3.2% on a developer-workspace fiducial.
+- The #490 regeneration itself is done and attributable — every other pin
+  touched in that pass reproduced its pre-fix value first
+  (`autolens_profiling` imaging cells to 1.3e-6, the developer jit cell
+  bit-for-bit, its RTU variant to 3.6e-10).
diff --git a/tasks/imaging_production_over_sampling.md b/tasks/imaging_production_over_sampling.md
new file mode 100644
index 0000000..91a3368
--- /dev/null
+++ b/tasks/imaging_production_over_sampling.md
@@ -0,0 +1,77 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autolens_profiling/imaging_production_over_sampling.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Profile imaging pixelizations at production over-sampling
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+Difficulty: large
+Autonomy: supervised
+Priority: normal
+Status: formalised
+Consequence: glance
+Witness: an A100 row in `results/runtime/imaging/delaunay/` records `over_sample_size_pixelization_rule` = {source_snr_cut 3.0, sub_size_above 4, sub_size_below 2} on Euclid data (today every A100 row records None)
+Review-minutes: 3
+Unattended: ready
+
+## Why
+
+A Euclid session compared a 1.5 s cold likelihood evaluation from a real Euclid
+fit against the autolens_profiling imaging headlines and found them not
+comparable. On checking (2026-09-27):
+
+- The JAX Delaunay runtime cell hard-codes `over_sample_size_pixelization=1`
+  (`scripts/imaging/likelihood_runtime/delaunay.py:244`). The headline A100 row
+  (`results/runtime/imaging/delaunay/delaunay_hpc_a100_fp64.json`, 65 ms per
+  eval, AdaptSplit) records no pixelization over-sampling rule, uses HST data
+  (0.05", 15 361 masked pixels) and the `[4, 2, 2]` light-profile bins.
+- Euclid production (`EUCLID_VIS_PIX` in `_production_config.py`, vis_pix stage,
+  job 342301) uses pixelization sub-size 4 above source S/N 3 and 2 below, with
+  `[4, 4, 2]` light-profile bins. The HST subhalo preset (`hst_source_pix_2`)
+  uses the same 4/2 pixelization rule.
+- The fixed-lens-light numba result (932 → 459 ms, HST Delaunay N1500, job
+  343311, `results/notes/fixed_lens_light_numba_2026_09.md`) was measured with
+  the older cell settings (60×1 MGE, no pixelization over-sampling rule
+  recorded). No fixed-light row exists at production settings.
+- The only production-preset numba row
+  (`results/runtime/imaging/delaunay_numba/delaunay_numba_likelihood_summary_hst_v2026.8.17.1.json`,
+  linear 30×2 MGE, 4/2 over-sampling) reads 1.46 s cold / 1.97 s warm, and its
+  witness against production job 342311 (0.52–0.81 s) is **FAIL**: about 2×
+  slower than production for an unexplained reason.
+
+The AdaptSplit vs ConstantSplit difference on the A100 is about 5 ms (65 vs 60
+ms) and is not the gap.
+
+## What
+
+1. Make the JAX imaging pixelization cells (runtime + breakdown, Delaunay
+   family, and rectangular where relevant) accept the production presets
+   (`--instrument euclid|hst`) so pixelization over-sampling follows the 4/2
+   S/N rule instead of a flat 1. Keep the legacy flat-1 setting selectable so
+   the historic rows can still be reproduced.
+2. Re-measure the A100 JAX headline at production over-sampling, on HST and on
+   **Euclid data with the `EUCLID_VIS_PIX` preset**, so the headline matches
+   what Euclid production runs.
+3. Add a fixed-lens-light numba sparse row at production settings (HST, and
+   Euclid if cheap), next to the linear-light production row, so both answers
+   to "how long does HST take, light linear vs fixed?" are production-matched.
+4. Investigate the `delaunay_numba` HST production-preset witness FAIL (cold
+   1.46 s vs job 342311's 0.52–0.81 s): is it the cell, the host, or the
+   reference?
+5. Update `results/README.md` and the campaign status note so headline
+   numbers state their over-sampling, preset and dataset, and older flat-1
+   rows are labelled as such.
+
+## Out of scope
+
+Library changes to the over-sampling code itself; the fitting library's
+summary-time evaluation timing (a separate question about the 1.5 s figure
+being one cold call in the main process).
+
+<!-- formalised by the Intake (Conception) Agent on 2026-09-27 from file:/tmp/claude-1000/-home-jammy-Code-PyAutoLabs/44f36318-0f57-4a47-b6a1-c66501ff840f/scratchpad/profiling_production_over_sampling.md -->
diff --git a/tasks/interferometer_decision_matrix_last_cell.md b/tasks/interferometer_decision_matrix_last_cell.md
new file mode 100644
index 0000000..aca78af
--- /dev/null
+++ b/tasks/interferometer_decision_matrix_last_cell.md
@@ -0,0 +1,66 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autolens_profiling/interferometer_decision_matrix_last_cell.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Interferometer decision matrix — fill the last cell (CPU rect 39² at alma_high r5.0) and flip the wiki row to shipped
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+Themes:
+- interferometer
+- numba-cpu
+Difficulty: small
+Autonomy: supervised
+Priority: normal
+Status: draft
+Consequence: glance
+Witness: `results/breakdown/interferometer/alma_high/pixelization_numba_hpc_ral_cpu_fp64_r5.0.json` is committed. Its gated arm has `configuration.inversion_path == InversionInterferometerSparseNumba`. Numba vs FFT agree to ≤ 0.5 nat, and CPU vs A100 (the 198.6 ms A100 row) agree to ≤ 1e-3 nat. The note's rect alma_high r5.0 row no longer reads "pending".
+Review-minutes: 4
+Unattended: needs-human
+Lane: local-dev
+Epic: interferometer-likelihood-campaign
+Filed: 2026-09-30
+
+## Context
+
+- Status: split out of `interferometer-decision-matrix` at close-out. Phase 4 shipped 13 of 14 cells in
+  `complete/2026/09/interferometer-decision-matrix.md` (autolens_profiling#358, merge `227b5c91`), and this
+  is what remains.
+- The campaign map is retired at `complete/archive/epics/interferometer_likelihood_campaign.md`.
+- The CPU rect 39² cell at alma_high r5.0 (RAL job **375978_3**) was still RUNNING at merge. The note
+  `results/notes/interferometer_likelihood_decision_matrix_2026_09.md` marks it "pending".
+
+## What
+
+1. **Pull.** Job 375978_3 writes `alma_high/pixelization_numba_hpc_ral_cpu_fp64_r5.0.{json,png}` into
+   `/mnt/ral/jnightin/autolens_profiling_wt/interferometer-decision-matrix/results/breakdown/interferometer/`.
+   Check `sacct -j 375978_3` reads COMPLETED and that the `.err` has 0 Tracebacks. Then copy both files into
+   `results/breakdown/interferometer/alma_high/` in a fresh task worktree off main.
+2. **Verify** before quoting a number:
+   - the gated arm's `configuration.inversion_path` is `InversionInterferometerSparseNumba`;
+   - numba vs NumPy FFT log-evidence agrees to ≤ 0.5 nat;
+   - CPU vs A100 agrees to ≤ 1e-3 nat against the existing A100 rect alma_high r5.0 row (198.6 ms);
+   - `source_revisions` match the private clone SHAs recorded in the note (Nerves 1ec1c82, Fit b13169e2,
+     Array 7a89e19a, Galaxy 4c834ced, Lens efd13c4c).
+3. **Fill the note.** Update the rect alma_high r5.0 row, the status line and the blocked-cells table in
+   `results/notes/interferometer_likelihood_decision_matrix_2026_09.md`. Recheck rule 4's 43–114× alma_high
+   range against the new row. Mark 375978_3 COMPLETED with its wall time in the RAL jobs line.
+4. **Flip `wiki/index.md`.** The interferometer campaign row should read shipped (phase 4 merged, all cells
+   filled). If the campaign page's Phase-4 section also says "pending", update it too.
+5. **Regenerate** `build_readme.py` and `build_dashboard.py` (and run each with `--check`), then run the repo lint
+   (check_wiki, check_results_layout, check_submits, ruff, pytest). Open one small PR.
+6. **Afterwards** (after merge, at close-out): remove the RAL worktree
+   `/mnt/ral/jnightin/autolens_profiling_wt/interferometer-decision-matrix` and the private library clone
+   `/mnt/ral/jnightin/PyAuto_branch/interferometer-decision-matrix`, but only once no job of this campaign
+   is running from them (`squeue -u jnightin`).
+
+## Done when
+
+- The JSON + PNG are committed, the note row is filled and verified, `wiki/index.md` reads shipped, the READMEs and
+  dashboard are regenerated, and lint is green on a merged PR.
+- The RAL worktree and private library clone are removed.
diff --git a/tasks/interferometer_fixed_mapper_curvature_preload.md b/tasks/interferometer_fixed_mapper_curvature_preload.md
new file mode 100644
index 0000000..435c768
--- /dev/null
+++ b/tasks/interferometer_fixed_mapper_curvature_preload.md
@@ -0,0 +1,77 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autolens_profiling/interferometer_fixed_mapper_curvature_preload.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Interferometer fixed-mapper searches: reuse the W~ curvature matrix across likelihood calls via `preloads.curvature_matrix`
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+- PyAutoLens
+- PyAutoArray
+Themes:
+- interferometer
+- pixelization
+- likelihood-profiling
+- jax-gpu
+Difficulty: medium
+Autonomy: supervised
+Priority: normal
+Status: draft
+Consequence: judge
+Witness: A note section lists which production interferometer search stages (SLaM source-pix and any regularization-only stage) have a mapper that is constant across the search, each with a code citation. `results/breakdown/interferometer/{alma,jvla}` gains a fixed-mapper row: library call with `preloads.curvature_matrix` injected vs without. The two agree on the figure of merit to ≤ 1e-9 nats, and the row gives the measured saving against the F row (36–42 % alma, 91–92 % jvla).
+Review-minutes: 12
+Unattended: needs-slicing
+Lane: local-dev
+Epic: interferometer-likelihood-campaign
+Filed: 2026-09-27
+
+Follow-up to autolens_profiling#320, lever 3
+(`results/notes/interferometer_mesh_a100_breakdown_2026_09.md`, "Ranked levers" §3).
+
+## Why
+
+On the A100 the sparse-path pixelized interferometer likelihood spends 36–42 % (alma), 66–73 %
+(alma_high) and 91–92 % (jvla) of the call building F = Aᵀ W~ A. At jvla Delaunay-1500 that is
+510 of 563 ms.
+
+- `InversionInterferometerSparse.curvature_matrix` already returns an injected
+  `preloads.curvature_matrix` untouched (PyAutoArray
+  `inversion/inversion/interferometer/sparse.py:162-163`).
+- The datacube case is shipped: `AnalysisInterferometer(shared_preloads=True)` builds F once
+  per evaluation and shares it across channels via `shared_state_from` (PyAutoLens
+  `interferometer/model/analysis.py:185-227`). That amortises F across channels **within** one
+  call.
+- Nothing reuses F **across** calls. When the mass is fixed and the image mesh is fixed, the
+  mapper (and so A, F and D) is constant for the whole search, and only H, the solve and the
+  log-dets change.
+- Bound: skipping F takes jvla Delaunay from 563 ms to ~53 ms and alma from 49.5 ms to
+  ~32 ms. This is a bound, not a measurement.
+
+## What
+
+1. **Census (read-only).** Which production interferometer stages have a constant mapper?
+   Check the SLaM source-pix stages with mass fixed from source_lp, regularization-only
+   searches, and the adapt-image / Hilbert image-mesh dependence (does the image mesh change
+   between calls in those stages?). Cite file:line for each. If no production stage qualifies,
+   stop here and record it.
+2. **Measure (harness-injected).** For each qualifying configuration, run the alma and jvla
+   Delaunay / rect cells (`scripts/interferometer/likelihood_breakdown/{delaunay,pixelization}.py`)
+   with F built once and passed as `aa.PreloadsInterferometer(curvature_matrix=F)` through
+   `fit_from(instance, preloads=...)`. Report per-call time vs the library row, the ≤ 1e-9 nats
+   figure-of-merit match, and `jit(vmap)` behaviour. A closed-over F constant must not bloat the
+   compiled program; record its size.
+3. **Design sketch only.** An opt-in on `AnalysisInterferometer` (e.g. cache F keyed on the
+   mapper identity, or a stage-level `fixed_mapper=True`) that reuses the `shared_preloads`
+   plumbing. The invariance contract must be explicit: the user or pipeline asserts the mapper
+   is fixed. Any library change is a separate prompt.
+
+## Not in scope
+
+Datacube channel sharing (already shipped). Mixed mapper + MGE inversions: the MGE columns
+change with the lens light, although the mapper-mapper block could still be reused; note it
+if the census finds it.
diff --git a/tasks/interferometer_nnls_memo_scattered_stream_guard.md b/tasks/interferometer_nnls_memo_scattered_stream_guard.md
new file mode 100644
index 0000000..885a1c2
--- /dev/null
+++ b/tasks/interferometer_nnls_memo_scattered_stream_guard.md
@@ -0,0 +1,71 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autolens_profiling/interferometer_nnls_memo_scattered_stream_guard.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# fnnls warm-start memo: stop it slowing scattered evaluation streams (interferometer CPU, autolens_profiling#332)
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+- PyAutoArray
+Themes:
+- interferometer
+- numba-cpu
+- nnls
+- likelihood-profiling
+Difficulty: medium
+Autonomy: supervised
+Priority: medium
+Status: draft
+Consequence: glance
+Witness: on a recorded real-sampler evaluation sequence (Nautilus, alma Delaunay-1500 and rect 39², NumPy numba route), the memo-on fnnls solve is no slower than memo-off on every phase of the run (early live points and late phase), while keeping >= 80 % of the local-walk gain measured in #332 (rect solve 0.13×, Delaunay 0.68×); figure of merit unchanged (Δ 0 nat).
+Review-minutes: 10
+Lane: local-dev
+Epic: interferometer-likelihood-campaign
+Filed: 2026-09-27
+
+## Context
+
+The NNLS passive-set memo is **on by default** (PyAutoArray `autoarray/config/general.yaml:12`,
+#498). The positive-only solve seeds from the previous evaluation's final passive set
+(`autoarray/inversion/inversion/nnls_memo.py`; used in `reconstruction_positive_only_from`,
+`autoarray/inversion/inversion/inversion_util.py:291`).
+
+autolens_profiling#332's lever arm (`levers.memo` in
+`results/breakdown/interferometer/{delaunay,pixelization}_numba_hpc_ral_cpu_fp64_levers.json`,
+alma r3.5, 16 instances per stream) measured:
+
+| Stream | Delaunay solve off → on | rect solve off → on |
+|---|---|---|
+| local walk (unit step 0.002) | 163 → 111 ms (0.68×) | 324 → 41 ms (**0.13×**) |
+| iid (central 20 % of each prior) | 164 → 356 ms (**2.17× slower**) | 329 → 356 ms (1.08× slower) |
+
+The solve is 17 % (Delaunay) / 22 % (rect) of the alma call, and 54 % / 76 % at sma. So the memo is
+worth −15 % of the rect call on a walk, but +15 % on the Delaunay call on an iid stream.
+
+The fallback guard (`inversion_util.py:598-630`) judges a seed only **after** the seeded solve has
+run. It drops the entry when `error_fraction > nnls_warm_start_error_tolerance ×
+dense_error_fraction`. The next solve restarts dense and re-seeds, so on a scattered stream about
+every other solve pays for a bad seed.
+
+## What
+
+1. Record the evaluation sequence (unit vectors) of a short Nautilus run on the alma interferometer
+   Delaunay / rect models. Replay it through the #332 harness's memo lever
+   (`_lever_memo`, `scripts/misc/likelihood_breakdown/interferometer_pixelized_numpy.py`), phase by
+   phase.
+2. If the iid-phase regression reproduces, prototype in-cell (no library edit) two guards and
+   measure both on the replay:
+   - a pre-solve seed check (e.g. KKT residual of the seed on the new system);
+   - a per-key back-off after k consecutive fallbacks.
+3. Report the imaging HST Delaunay control too (the #498 fiducial), so the fix does not undo the
+   imaging gain.
+4. If a guard wins, draft the PyAutoArray change.
+
+## Out of scope
+
+The JAX solvers (memo is NumPy-only).
diff --git a/tasks/interferometer_streaming_scaling.md b/tasks/interferometer_streaming_scaling.md
new file mode 100644
index 0000000..150db42
--- /dev/null
+++ b/tasks/interferometer_streaming_scaling.md
@@ -0,0 +1,48 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autolens_profiling/interferometer_streaming_scaling.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Campaign: interferometer streaming (array-free) vs in-memory — memory and time scaling to 2e8 visibilities
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+Themes:
+- interferometer
+- sparse-operator
+- memory
+Difficulty: small
+Autonomy: supervised
+Priority: high
+Status: draft
+Consequence: glance
+Witness: `scripts/interferometer/streaming_scaling/` holds the campaign cells (accumulator time scaling vs N_vis and chunk size; in-memory `apply_sparse_operator` pushed to failure under a memory cap; streaming at the failing sizes; log_evidence parity), each writing a versioned JSON row under `results/streaming_scaling/` through `_profile_cli.py`; `build_readme.py --check`, `check_results_layout.py`, `check_wiki.py` and `ruff` pass; `wiki/campaigns/interferometer_streaming.md` (from `_template.md`) records the table, the 2e8 extrapolation and the go/no-go on the streaming epic's remaining phases; the dashboard is regenerated.
+Review-minutes: 5
+Unattended: ready
+Epic: streaming-visibilities
+
+Source: the 2026-09-30 go/no-go question on Discussion https://github.com/orgs/PyAutoLabs/discussions/13 (HRSAstro's 2e8-sample ALMA cube). Phases 1-2 of the streaming epic are merged (PyAutoArray#593, PyAutoGalaxy#639, PyAutoLens#758). A scratchpad run of this benchmark was made in the CLI session on 2026-09-30 (`bench_stream.py`); this campaign ports it into the repo so the evidence is versioned and re-runnable per release.
+
+## Scratchpad result to reproduce (2026-09-30)
+
+In-memory `apply_sparse_operator` OOMs at 1e6 visibilities under a 10 GB cap (3.1 GB allocation in `nufft_precision_operator_via_nufft_from`; 2.1 GB peak already at 1e5). Streaming at chunk 65536: 16 / 52 / 174 / 539 s at 1e6 / 4e6 / 1.6e7 / 5e7 (≈11 s per 1e6 vis, linear), peak RSS 1.5–1.8 GB flat; chunk 4096 is 5× slower (per-chunk fixed cost: nufft spread 46 %, JAX recompiles ~73 across 40 chunks, device→host copies). 2e8 ⇒ ~37 min, ~1.8 GB. Verdict GO; the table is on the epic ledger.
+
+## Why
+
+The array-free dataset only earns its maintenance cost (a second dataset kind every interferometer code path branches on) if it makes a real difference at the visibility counts real data reach. In-house datasets are ≤1.1e5 visibilities; the discussion's case is 2e8. The decision on phases 3-5 (visualizer, non-linear light profiles, cubes) rests on measured crossover memory and on the accumulator being linear and fast enough to reach 2e8.
+
+## What
+
+1. Cells under `scripts/interferometer/streaming_scaling/` (dataset-first, task-second layout; use `_profile_cli.py` for JSON/CLI; synthetic seeded per-chunk visibilities generated in memory, 400-px circular mask at 0.05"/pix, `TransformerNUFFT`; each measurement in a fresh child process with a `RLIMIT_AS` cap and per-child timeout):
+   - `accumulate.py` — `Interferometer.from_stream` wall time and peak RSS vs N_vis ∈ {1e6, 4e6, 1.6e7, 5e7} × chunk ∈ {4096, 65536}; seconds per 1e6 vis; linearity.
+   - `in_memory.py` — `Interferometer(...).apply_sparse_operator()` peak RSS + wall at N_vis ∈ {1e6, 4e6, 1.6e7, 5e7, 1e8}; the first N that fails under the cap is a result.
+   - `parity.py` — log_evidence of a 20×20 rectangular sparse inversion, streamed vs in-memory at 4e6.
+2. `results/streaming_scaling/` rows + PNG (RSS and wall vs N_vis, both paths) following `check_results_layout.py`; README dashboard via `build_readme.py`; `build_dashboard.py` regenerated.
+3. `wiki/campaigns/interferometer_streaming.md` from `_template.md`: table, 2e8 extrapolation (in-memory RSS ≈ 96 B/vis + temporaries vs streaming RSS; accumulation time at the best chunk), a cProfile top-10 of one chunk if any rate exceeds 5 s per 1e6 vis, and the go/no-go for the epic's phases 3-5. Link from `wiki/index.md`.
+4. A100 rows (RAL) are optional follow-ups; CPU rows decide the memory question.
+
+Parallel claims: autolens_profiling is claimed by `raw-pdip-forward-polish` (`interferometer-decision-matrix` shipped 2026-09-30, `complete/2026/09/interferometer-decision-matrix.md`); this campaign adds a new task folder and results folder only (shared files: `wiki/index.md` rows + generated README/dashboard).
diff --git a/tasks/interferometer_w_tilde_fft_size_levers.md b/tasks/interferometer_w_tilde_fft_size_levers.md
new file mode 100644
index 0000000..e009303
--- /dev/null
+++ b/tasks/interferometer_w_tilde_fft_size_levers.md
@@ -0,0 +1,83 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autolens_profiling/interferometer_w_tilde_fft_size_levers.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Interferometer W~ curvature matrix is FFT-bound on the mask extent: pruned padded FFT and real-space pixel scale on the A100
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+- PyAutoArray
+Themes:
+- interferometer
+- pixelization
+- likelihood-profiling
+- jax-gpu
+Difficulty: medium
+Autonomy: supervised
+Priority: normal
+Status: draft
+Consequence: glance
+Witness: `results/breakdown/interferometer/jvla/delaunay_hpc_a100_fp64_fft_prune.json` has a pruned-vs-library F row at alma, alma_high and jvla. The pruned F agrees with the library F to max rel diff ≤ 1e-12 and the figure of merit to ≤ 1e-6 nats. The file also has an F-vs-real-space-pixel-scale row at jvla (0.01 / 0.015 / 0.02 / 0.025″) with the log-evidence shift against the 0.01″ reference.
+Review-minutes: 10
+Unattended: ready
+Lane: local-dev
+Epic: interferometer-likelihood-campaign
+Filed: 2026-09-27
+
+Follow-up to autolens_profiling#320, lever 2
+(`results/notes/interferometer_mesh_a100_breakdown_2026_09.md`, "Ranked levers" §2).
+
+## Why
+
+On the A100 the W~ curvature matrix F = Aᵀ W~ A is 36–42 % of the sparse-path pixelized
+likelihood at alma, 66–73 % at alma_high and 91–92 % at jvla (510 of 563 ms, Delaunay-1500).
+Its cost is set by the mask extent M = y·x (70² / 140² / 280² / 700² at sma → jvla), not by
+N_vis.
+
+- The block-size sweep is flat (alma B = 32–512: 17.8–19.3 ms).
+- The kernel split puts the FFT apply at 71–85 % of F.
+- So F is **FFT-bound** on the (B, 2y, 2x) `rfft2` / `irfft2` pair.
+
+The extent is already the mask bounding box (PyAutoArray `mask/mask_2d.py:746-776`
+`extent_index_for_masked_pixel`), and the (2y, 2x) pad is the minimum for a linear
+correlation. Only two things are left to change: how much of the padded transform is computed,
+and how big the extent is.
+
+## What
+
+1. **Pruned padded transform (measurement first, harness-only).**
+   `InterferometerSparseOperator.apply_operator` (PyAutoArray
+   `inversion/inversion/interferometer/inversion_interferometer_util.py:1099-1151`) pads each
+   (B, y, x) block to (B, 2y, 2x) and runs full 2-D `jnp.fft.rfft2` / `irfft2` (`:1145-1150`).
+   Only the top-left y × x quadrant of the input is non-zero, and only the top-left y × x of the
+   output is kept (`:1150`).
+   - Write a harness copy of `curvature_matrix_diag_from` (`:1171-1253`) whose apply is
+     separable: `rfft` along x on the y non-zero rows (x padded to 2x), zero-pad y, `fft` along
+     y, multiply by `Khat`, `ifft` along y and keep y rows, `irfft` along x with `n = 2x`,
+     keep x.
+   - That removes up to ~25 % of the FFT work. The bound is ≤ ~15–20 % of F (~90 ms at jvla,
+     ~3 ms at alma).
+   - Measure it against the library F inside the same fused jit at alma / alma_high / jvla,
+     Delaunay and rect. One cuFFT 2-D call may beat four 1-D calls, so record a no-go honestly.
+2. **Real-space pixel scale at jvla (measurement, science-gated).** jvla at 0.01″ is a 700²
+   extent. Re-run the jvla Delaunay-1500 cell at real-space pixel scales 0.015 / 0.02 / 0.025″
+   (mask radius fixed at 3.5″; `instruments/interferometer.py` jvla entry). Record F, the full
+   call and the log-evidence / reconstruction shift against 0.01″. Predicted: 2× coarser gives
+   ~4× smaller M, putting F near the alma_high row (~67 ms). Whether the science tolerates it is
+   the question, not the speed. The mask-radius half (2.0 / 3.5 / 5.0″) is in task 3/3's
+   decision matrix on CPU (`complete/archive/epics/interferometer_likelihood_campaign.md`, retired 2026-09-30; the matrix is `results/notes/interferometer_likelihood_decision_matrix_2026_09.md`);
+   do not duplicate it.
+3. Verdict in a note section. If (1) wins by ≥ 10 % of F at jvla with a ≤ 1e-12 F match, file
+   the PyAutoArray change (the NumPy branch `:1155-1169` gets the same pruning via
+   `scipy.fft`). If (2) holds the 0.5-nat bar at a coarser scale, file a workspace docs prompt
+   for interferometer real-space grid guidance.
+
+## Not in scope
+
+The fp32 FFT arm (slower on the A100 and fails the bar at jvla, #320), block size (flat), and
+mixed precision (does not reach F).
diff --git a/tasks/jax_compile_probe_needs_own_cell_builder.md b/tasks/jax_compile_probe_needs_own_cell_builder.md
new file mode 100644
index 0000000..4fd5f09
--- /dev/null
+++ b/tasks/jax_compile_probe_needs_own_cell_builder.md
@@ -0,0 +1,42 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/maintenance/autolens_profiling/jax_compile_probe_needs_own_cell_builder.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# jax_compile/probe.py lost its cell builder with the searches tier — give profiling its own
+
+Type: maintenance
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+Themes:
+- jax
+- compile
+- hygiene
+Difficulty: medium
+Autonomy: supervised
+Priority: normal
+Status: draft
+Consequence: judge
+Witness: `python scripts/misc/jax_compile/probe.py --cell <one pinned cell>` runs end to end and reproduces that cell's pinned warm-compile record within its tolerance
+Review-minutes: 20
+Unattended: ready
+Filed: 2026-09-10
+
+`scripts/misc/jax_compile/probe.py` built its cells through `searches._setup.build_for_cell`
+(a 2,159-line module of the retired inference programme). autolens_profiling#245 removed
+that tier (archived as PyAutoGut ref `archive/condemned/autolens-profiling/inference-programme`
+@ `c8b60580`); `build_objective` now raises `NotImplementedError` naming the ref, so the
+probe cannot run. Its pinned warm-compile records and the compile dashboard are intact.
+
+Give the probe a builder this repo owns: the runtime/breakdown scripts already construct
+every pinned cell (`_production_config.preset_for` + the per-cell model code in
+`scripts/imaging/likelihood_runtime/*.py`), so factor the cell construction they share
+into a small `scripts/misc/cells.py` (or reuse an existing helper if one exists — survey
+first) and point `build_objective` at it. Do not restore anything from the archive ref.
+Re-run the probe on every pinned cell and confirm the records still match.
+
+Filed at the close of #245 (scrap-inference-programme); see
+`complete/2026/09/autolens-inference-birth.md` for the epic.
diff --git a/tasks/mass_field_flat_adoption_science_repos.md b/tasks/mass_field_flat_adoption_science_repos.md
new file mode 100644
index 0000000..9af4088
--- /dev/null
+++ b/tasks/mass_field_flat_adoption_science_repos.md
@@ -0,0 +1,120 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/maintenance/autolens_profiling/mass_field_flat_adoption_science_repos.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Remaining code consumers move to the flat `fields=` form: profiling, inference, JOSS benchmarks, Reduce prototypes
+
+**2026-09-19 update:** This four-repo survey is now historical context. PyAutoReduce#77 and both autolens_inference consumers (#8 and #10) merged. The user confirmed the deleted local `autolens_jax_joss` checkout is outside this sweep. The 47 live non-chaining profiling calls and all 5 staged resume calls shipped in `complete/2026/09/mass-field-profiling-live.md` and `complete/2026/09/mass-field-pipeline-resume.md`; 5 inline witnesses remain historical. Do not issue this broad prompt as a single task.
+
+Type: maintenance
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+- autolens_inference
+- autolens_jax_joss
+- PyAutoReduce
+Themes:
+- cluster
+- profiling
+- hygiene
+Difficulty: large
+Autonomy: supervised
+Priority: normal
+Status: draft
+Consequence: glance
+Witness: an AST re-walk (never a grep — memory `ASTwitness`) over each repo's live script set reports zero `af.Model(al.Galaxy, ..., shear=...)` and zero `al.Galaxy(..., shear=...)` model builders outside the named exclusions, and every migrated builder composes `fields=field` with `field = af.Model(al.MassField, redshift=..., shear=af.Model(al.mp.ExternalShear))`; one runnable script per repo executes end to end; `autolens_profiling` lint (`build_readme.py --check`) passes with regenerated dashboards.
+Review-minutes: 10
+Unattended: needs-slicing
+Epic: mass-field
+Superseded-by: `complete/2026/09/mass-field-profiling-live.md`, `complete/2026/09/mass-field-pipeline-resume.md`, and `complete/2026/09/mass-field-inference-slam.md`; helper merged in PyAutoGalaxy#625 / PyAutoLens#745 and consumed from current source `main`
+Filed: 2026-09-18
+
+## The trap, stated first
+
+These four repos returned **zero `fields=` hits**. That is **not** because they are
+clean — it is because they were **never migrated**. They sit at the **pre-epic,
+galaxy-attached** form: `af.Model(al.Galaxy, ..., shear=af.Model(al.mp.ExternalShear))`.
+Reporting "0 hits, nothing to do" would be exactly wrong, and is the reading this
+prompt exists to prevent. Zero-hit ≠ done.
+
+Target idiom (the flat form shipped by epic phase 6):
+
+```python
+field = af.Model(al.MassField, redshift=<lens z>, shear=af.Model(al.mp.ExternalShear))
+model = af.Collection(galaxies=af.Collection(lens=lens, source=source), fields=field)
+# prior paths read fields.shear.gamma_1, not fields.field.shear.gamma_1
+```
+
+## Dependency on the chaining helper
+
+`al.util.chaining.mass_from` / `chaining_util.mass_from` takes **no `fields`
+argument**, so a chained stage that carries galaxies forward and omits `fields=`
+**silently drops the external field** — no error, no log line. That is why this prompt
+is `Blocked-by` the helper prompt. Do not migrate any staged pipeline here by hand
+before the helper lands; a text replace across a chained file is the documented way to
+produce a wrong model that still runs.
+
+## Per repo
+
+### `autolens_profiling` — 52 live runnable model builders under `scripts/`
+
+- **No archive directory exists** in this repo. Experiment snapshots are *inline*,
+  which means the usual "everything outside `archive/` is live" rule does not apply.
+  **EXCLUDE from any rewrite** — these are recorded measurements, not live code:
+  - `fixed_light_numba_s4_witness.py`
+  - `*_levers_l{2,3}_witness.py`
+  - `fixed_light_draws.py`
+  - `fixed_light_trace.py`
+  - `results/hazards/component/profile_registry_coverage.json` (generated tripwire)
+- **READMEs are generated.** Lint runs `build_readme.py --check` (memory `READMEgen`),
+  so anything that touches `results/` rows needs the regenerated dashboards in the same
+  PR.
+- **Claim conflict:** the repo is currently claimed by `fixed-light-numba-s4b`.
+  `hst-gpu-residue-p2` completed in
+  `complete/2026/09/hst-gpu-residue-p2.md`; re-survey the remaining live claim before
+  starting and use a fresh worktree.
+
+### `autolens_inference` — 3 live files, one of them the risk
+
+- `scripts/misc/simulators/imaging.py`, `scripts/misc/simulators/interferometer.py` —
+  live simulators with galaxy-attached `ExternalShear` instances.
+- `scripts/misc/slam/_runner.py` — **a LIVE staged SLaM pipeline**, chaining
+  `Isothermal + ExternalShear` into `PowerLaw` via
+  `al.util.chaining.mass_from(..., unfix_mass_centre=True)` (module docstring lines
+  15-20, builder at lines 713-742). **This file is precisely why the helper prompt
+  blocks this one.** Migrate it last, with a per-stage witness that the field survives
+  into every stage's model.
+
+### `autolens_jax_joss` — 7 benchmark files
+
+`benchmarks/{group,imaging,imaging_and_interferometer,imaging_and_point_source,
+interferometer,multi_band,point_source}.py`. All galaxy-attached, all live runnable
+benchmarks, no open PRs. Note this repo is **absent from `repos.yaml`** — see
+`draft/maintenance/pyautomind/autolens_jax_joss_manifest_gap.md`; adding it there is
+independent of this migration but should land first so the repo is visible to the next
+manifest-driven sweep.
+
+**Update 2026-09-18 (PyAutoBrain#389):** this repo's **local checkout was removed** as
+dead weight. The 7 benchmark files still exist on GitHub at `6bce65e` and still need
+the flat-`fields=` migration, but they are **no longer locally sweepable** — this
+member now requires a fresh clone, so sequence it accordingly (or drop it from the
+local sweep and raise it as a PR against the remote). The remote is deliberately
+retained: it is the citable artefact behind the JOSS paper's benchmark claims.
+
+### `PyAutoReduce/prototypes` — 2 files
+
+Smallest of the four; prototypes, no chaining.
+
+## Sequencing
+
+Four repos, so `Unattended: needs-slicing` — this is a slice-per-repo task, not one
+sitting. Suggested order: `PyAutoReduce` (2 files) → `autolens_jax_joss` (7) →
+`autolens_profiling` (52, needs the claim waiver and the README regeneration) →
+`autolens_inference` (3, the SLaM runner last, after the helper). Each repo is its own
+PR; the `mass-field` epic table should gain a row once the human rules on scope.
+
+Filed 2026-09-18 from the flat-`fields=` adoption sweep's follow-up audit
+(issue autolens_workspace#561).
diff --git a/tasks/mge_likelihood_breakdown_steps_are_cumulative_an.md b/tasks/mge_likelihood_breakdown_steps_are_cumulative_an.md
new file mode 100644
index 0000000..11dc8c0
--- /dev/null
+++ b/tasks/mge_likelihood_breakdown_steps_are_cumulative_an.md
@@ -0,0 +1,38 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/bug/workspaces/mge_likelihood_breakdown_steps_are_cumulative_an.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# MGE likelihood_breakdown steps are cumulative and `linear_gaussians` is reported as 0
+
+Type: bug
+Target: workspaces
+Repos:
+- autolens_profiling
+Difficulty: medium
+Autonomy: safe
+Priority: low
+Memory: wiki/galaxies/sources/massive-ellipticals.md; wiki/lensing/sources/dark-matter-substructure.md; reading-queue.md
+Status: formalised
+Consequence: glance
+Witness: After the fix, `likelihood_breakdown/mge.py` on the HST cell writes a JSON whose step timings sum to within 15% of the fused `likelihood_runtime/mge.py` single-call time (currently 40 ms vs 29.6 ms), and `configuration.linear_gaussians` equals the number of linear Gaussians inside the Basis objects (40 for the profiling cell), in both breakdown and runtime JSONs.
+Review-minutes: 3
+Unattended: ready
+
+# MGE likelihood_breakdown steps are cumulative and `linear_gaussians` is reported as 0
+
+Type: bug
+Priority: low
+Target: autolens_profiling
+Witness: After the fix, `likelihood_breakdown/mge.py` on the HST cell writes a JSON whose step timings sum to within 15% of the fused `likelihood_runtime/mge.py` single-call time (currently 40 ms vs 29.6 ms), and `configuration.linear_gaussians` equals the number of linear Gaussians inside the Basis objects (40 for the profiling cell), in both breakdown and runtime JSONs.
+
+
+Found in the 2026-09-24 MGE likelihood audit.
+
+1. `scripts/imaging/likelihood_breakdown/mge.py`: step 2 (`mapping_matrix_from_params`, ~:418) re-runs the ray-trace and step 3 (`blurred_mm_from_params`, ~:462) rebuilds steps 1-2, so the per-step timings are cumulative and `total_step_by_step` double-counts. The published `mge_breakdown_hst_v2026.8.17.1.json` says 40.3 ms; the fused runtime is ~29.6 ms, and the real step-sum is ~29 ms (blurred 25.6 + D/F/NNLS/image/chi2 3.4). The A100 `mge_hpc_a100_fd64.json` total 7.8 ms is really ~6.5 ms, which makes NNLS ~44% of an A100 evaluation, not 36%. Check whether the other breakdown scripts (pixelization.py, delaunay*.py, fixed_light*.py) share the pattern before fixing only mge.py.
+2. `tracer.cls_list_from(cls=al.lp_linear.LightProfileLinear)` (breakdown/mge.py ~:311, likelihood_runtime/mge.py ~:332) does not look inside an MGE `Basis`, so both scripts write `configuration.linear_gaussians: 0`; the real counts are 40 (profiling cell). Count the linear profiles inside each Basis.
+3. Fix both, regenerate the affected result JSONs/PNGs and the README dashboards (`build_readme.py --check` gates lint), and note in the JSON/README that pre-fix breakdown totals are cumulative and not comparable.
+
+<!-- formalised by the Intake (Conception) Agent on 2026-09-24 from file:/tmp/claude-1000/-home-jammy-Code-PyAutoLabs/4c46534e-38b6-47fd-9021-8040da7c7d92/scratchpad/mge_audit/prompt_breakdown.md -->
diff --git a/tasks/mge_nnls_fix_pyautoarray_571_slam_60.md b/tasks/mge_nnls_fix_pyautoarray_571_slam_60.md
new file mode 100644
index 0000000..6fff64d
--- /dev/null
+++ b/tasks/mge_nnls_fix_pyautoarray_571_slam_60.md
@@ -0,0 +1,57 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autoarray/mge_nnls_fix_pyautoarray_571_slam_60.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Linear-solver programme phase 3: GPU/vmap/A100 timing and parity rows for the solver corpus
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+Difficulty: too-large
+Autonomy: supervised
+Priority: medium
+Memory: wiki/lensing/sources/dark-matter-substructure.md; reading-queue.md; wiki/lensing/sources/lens-modeling-methods.md
+Status: formalised
+Filed: 2026-09-24, retargeted: 2026-09-30 (epic linear-solver-programme)
+Blocked-by: a release shipping the phase-2 fix (PyAutoArray#595, merged 2026-09-30 pending release; record complete/2026/09/raw-pdip-forward-polish.md) AND the RAL mirror synced via HPCPullPyAuto
+Witness: `scripts/lens/solver/timing.py` rows plus `accuracy.py --device gpu` rows on the phase-1 corpus (`results/lens/solver/corpus/`: slam_fixture_571, slam48_hst, slam_spread_hst, euclid_vis_lp — 81 systems) record, for the SLaM source_lp[1] 60-column model and the captured euclid system, single / vmap16 / vmap50 per-evaluation cost on the RTX 2060 and RAL A100 fp64 for jacobi vs raw (released phase-2 solver), the NNLS share of each batch, and GPU-vs-CPU parity (max |dlogL| and amplitude/`flux_inactive_rel` agreement) on the 48 near-truth vectors — appended to `results/notes/linear_solver_accuracy_2026_09.md` and `wiki/campaigns/linear_solver_accuracy.md`.
+Review-minutes: 3
+Consequence: glance
+Unattended: needs-slicing
+
+Epic `linear-solver-programme`, phase 3. Contract: `complete/2026/09/linear-solver-accuracy-study.md` (phase-1 record with the original programme prompt folded in).
+
+This prompt was filed on 2026-09-24 as a standalone PyAutoArray#571 follow-up (SLaM 60-column GPU
+timing and parity, single and vmap). The original ask below now lives in the phase-1
+`scripts/lens/solver/` package and its corpus rather than in a fresh
+`scripts/imaging/hazards/` script: the SLaM model recipe, the 48 near-truth vectors and the
+euclid capture are already corpus systems, so phase 3 adds GPU/A100 cells to that package and
+appends rows to the shared ledger. Phase 1 found the released raw stop leaves spurious
+amplitude on reference-inactive columns while logL barely moves, so GPU parity must be scored
+with `flux_inactive_rel` (and amplitude agreement), not logL alone.
+
+## Context
+
+Context: the 2026-09-24 audit measured only CPU timing for the fix (neutral: 27.3 -> 25.8 ms single, 31.3 -> 30.3 ms/eval vmap16). On GPU the pre-fix behaviour made every Nautilus vmap batch run all 50 PDIP iterations whenever one lane failed (NNLS ~28% of a 2060 batch, ~44% of a single A100 evaluation), so the fix should cut batch cost; that is unmeasured. GPU parity of the raw-forward PDIP is also unmeasured (the audit saw CPU/GPU/vmap divergence on the failing vectors pre-fix).
+
+## Ask
+
+1. Add a new cell to `scripts/lens/solver/timing.py` following the package's `_driver`
+   pattern, loading the SLaM source_lp[1] 60-column system and the euclid capture from
+   `results/lens/solver/corpus/`; time single, vmap16 and vmap50 per-evaluation cost on the
+   RTX 2060 and RAL A100, fp64, with `nnls_preconditioning_no_mapper` forced to jacobi and to
+   raw (the released phase-2 solver), interleaved minima.
+2. Record NNLS share per config via the existing solve ablation pattern (unconstrained solve
+   swap) or `stats["iterations"]`.
+3. GPU parity: run `accuracy.py --device gpu` over the corpus (at minimum the 48 near-truth
+   vectors); report max |logL_gpu - logL_cpu|, amplitude agreement and `flux_inactive_rel`
+   against the CPU fnnls reference, and iterations per lane.
+4. Append the rows and a verdict to `results/notes/linear_solver_accuracy_2026_09.md` and
+   `wiki/campaigns/linear_solver_accuracy.md` (regenerate README dashboards,
+   `build_readme.py --check`); route any regression to /intake as a bug.
+
+<!-- formalised by the Intake (Conception) Agent on 2026-09-24; retargeted into epic linear-solver-programme phase 3 on 2026-09-30 -->
diff --git a/tasks/numba_breakdown_harness_memo_blind.md b/tasks/numba_breakdown_harness_memo_blind.md
new file mode 100644
index 0000000..29b7b40
--- /dev/null
+++ b/tasks/numba_breakdown_harness_memo_blind.md
@@ -0,0 +1,51 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/feature/autolens_profiling/numba_breakdown_harness_memo_blind.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Numba breakdown harness: perturb the instance so the operated-matrix memo cannot hide a step
+
+Type: feature
+Target: autolens_profiling
+Repos:
+- @autolens_profiling
+Themes:
+- numba-cpu
+- profiling
+Difficulty: small
+Autonomy: safe
+Priority: medium
+Status: formalised
+Consequence: glance
+Witness: The `pixelization_numba` and `delaunay_numba` breakdown/runtime harnesses perturb the instance per repeat (or run with `AUTOARRAY_NUMBA_OPERATED_MEMO=0`) and record which in the results JSON `configuration`, keep the pinned log-likelihood check on the unperturbed instance, and the re-baselined euclid + hst rows show the operated-matrix step at its un-memoised cost with the regime change noted in the results README.
+Review-minutes: 3
+Unattended: ready
+Filed: 2026-08-27
+
+## Context (found while measuring PyAutoArray#496, 2026-08-27)
+
+Was a member of the `numba-cpu-likelihood` epic, retired 2026-09-02 to
+`complete/archive/epics/numba-cpu-likelihood.md`; it stands alone now.
+
+`scripts/imaging/likelihood_breakdown/pixelization_numba.py` and the
+`likelihood_runtime` sibling time `n_repeats=10` evaluations of one fixed
+`instance`. Since the cross-evaluation memo in PyAutoArray
+`imaging_numba/sparse.py` (sha256 of the pickled linear func), every repeat
+after warm-up hits the memo, so "MGE operated mapping matrix (60 funcs)" reads
+~0.003 s (euclid) / ~0.01 s (hst) — and `direct_log_likelihood_function_per_call`
+is contaminated the same way. Batching the convolution (a 3.7-5.4x win on that
+step, measured with `AUTOARRAY_NUMBA_OPERATED_MEMO=0`) showed A/B = 1.04 in the
+harness. Real modelling perturbs the MGE parameters every evaluation, so the
+memo never hits there.
+
+## Goal
+
+- Per repeat, perturb the instance (e.g. shift the lens-light Gaussian centres
+  by 1e-3·k) OR set `AUTOARRAY_NUMBA_OPERATED_MEMO=0` for the numba cells, so
+  the step is measured un-memoised; record which in the results JSON
+  `configuration`.
+- Keep the pinned log-likelihood check on the unperturbed instance.
+- Re-baseline the `pixelization_numba` and `delaunay_numba` breakdown/runtime
+  results (euclid + hst) and note the regime change in the results README.
diff --git a/tasks/pair_jax_xla_env_vars_with_measured.md b/tasks/pair_jax_xla_env_vars_with_measured.md
new file mode 100644
index 0000000..d66efe2
--- /dev/null
+++ b/tasks/pair_jax_xla_env_vars_with_measured.md
@@ -0,0 +1,133 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autonerves/pair_jax_xla_env_vars_with_measured.md
+Migrated-at: 2026-10-03
+
+Original task retained verbatim below. Current scheduling state: `campaigns.yaml`. Historical paths: `migration.yaml`.
+
+---
+
+# Pair JAX/XLA env vars with measured compile and run times, per backend
+
+Type: research
+Target: PyAutoNerves
+Repos:
+- autogalaxy_workspace_test
+- autolens_workspace_test
+- PyAutoNerves
+- workspaces
+Themes:
+- jax-compile
+- profiling
+Difficulty: too-large
+Autonomy: supervised
+Priority: high
+Status: formalised
+Consequence: judge
+Review-minutes: 25
+Unattended: needs-slicing
+
+# Pair JAX/XLA env vars with measured compile and run times, per backend
+
+Type: research
+Target: autonerves
+Repos:
+- @PyAutoNerves
+- @autolens_workspace_test
+- @autogalaxy_workspace_test
+Difficulty: too-large
+Autonomy: supervised
+Priority: high
+
+## Why this exists
+
+`complete/2026/08/xla-cpu-eigen-pool-deadlock.md` root-caused a real hang:
+`xla::cpu::FftThunk::Execute` runs on an Eigen intra-op pool worker and hands
+ducc0 that same pool, which fans the transform back into it and deadlocks. The
+avoidance is `XLA_FLAGS=--xla_cpu_multi_thread_eigen=false`, which costs ~15%.
+
+Two things that campaign exposed, neither of which it fixed:
+
+1. **The flag is CI-only.** It lives in `config/build/profile_{smoke,release}.yaml`
+   of the two test workspaces. No library sets it — verified against a fully
+   installed stack: zero hits for `xla_cpu_multi_thread_eigen` anywhere in
+   site-packages. So a **real science run on CPU has neither the protection nor
+   the 15%**, while CI pays the 15% on every JAX script. Both halves of that are
+   chosen by accident rather than by measurement.
+2. **Nobody has measured which JAX/XLA env vars actually matter.** `autonerves/
+   jax_wrapper.py` already manages `XLA_FLAGS` — it sets
+   `--xla_disable_hlo_passes=constant_folding` and `--xla_gpu_autotune_level=0`
+   — so there is a place a measured default belongs. Those two were set for
+   reasons that were true when written; neither has a current number attached,
+   and the eigen flag was added to CI without one either.
+
+## The question
+
+For each backend (CPU and GPU) and each candidate env var, what does it do to
+**compile time** and to **execution time**, separately? Those two are different
+budgets — this stack has repeatedly optimised one while paying in the other, and
+the epic that produced this task spent a month with a hang misnamed a "compile
+stall" precisely because they were not measured apart.
+
+## Scope
+
+Sweep candidate env vars across representative JAX likelihood scripts, record
+compile and run time separately per backend, and produce a **recommended default
+set per backend** with the measurement behind each entry. Candidates at least:
+
+- `--xla_cpu_multi_thread_eigen` (the incident flag; CPU)
+- `--xla_disable_hlo_passes=constant_folding` (already set; unmeasured)
+- `--xla_gpu_autotune_level` (already set to 0; unmeasured)
+- `JAX_COMPILATION_CACHE_DIR` / the persistent cache (on by default for a while
+  now; its effect on a warm vs cold runner is unquantified)
+- `JAX_ENABLE_X64` (correctness-bearing, but the cost is unrecorded)
+- thread-pool sizing knobs, and process CPU affinity, which the incident showed
+  changes behaviour even when no env var does
+
+## The constraint that shapes the plan
+
+**GPU measurement cannot run in GitHub Actions** — there are no GPU runners. The
+GPU leg has to go through the RAL HPC `gpu` partition via the project's
+`hpc/sync push-submit gpu` path, which is a different harness, a different
+cadence and a different results-ingest from the CPU leg. Plan the two legs as
+separate phases rather than assuming one sweep covers both; a design that only
+works on GitHub runners silently answers half the question.
+
+## Suggested phasing
+
+Sized `too-large` deliberately — split at start_dev.
+
+1. **CPU sweep harness + baseline.** Extend or sibling the `retime.py` harness,
+   which already has `--arm` for env overlays dealt round-robin within one
+   dispatch (added by the incident above). Record compile and run time
+   separately. Deliverable: a per-var CPU table.
+2. **GPU sweep** via `hpc/sync push-submit gpu`, same measurements, GPU-relevant
+   vars. Deliverable: a per-var GPU table.
+3. **Decide and land the defaults** in `autonerves/jax_wrapper.py`, per backend,
+   each with its number. Includes the decision on whether the eigen flag becomes
+   a user-facing CPU default — which is the science-run exposure question — and
+   whether CI should keep paying 15% on scripts that never trip the deadlock.
+
+## Do not re-measure
+
+Established by the incident and in its record; re-running these is waste:
+
+- The eigen flag's *effect on the deadlock* is settled: standalone reproducer
+  hangs 0/8 by default and 8/8 with the flag, Fisher p = 0.000155; on the real
+  script, pool of 4 hung 5/6 and pool of 1 passed 6/6.
+- Its cost on `imaging/jax_likelihood/mge_group.py` is ~15% (62.3s vs 54.4s).
+- The persistent compilation cache is **not** the deadlock's cause (6/6 hangs
+  with it disabled) — that says nothing about its *timing* effect, which is what
+  this task should measure.
+- There is no CFS quota on `ubuntu-latest`, so pool mis-sizing against a cgroup
+  quota is not a live variable there.
+
+## Acceptance
+
+- A per-backend table: env var x {compile time, run time}, with the script set
+  and repeat count stated, and a re-measured control (the incident's own lesson:
+  timings and hang rates wander, so before/after is not evidence — ABAB is).
+- A recommended default set per backend, each entry carrying its number.
+- An explicit decision on the CPU eigen flag for science runs: made a library
+  default, left CI-only, or documented for users — with the reason and the cost.
+- Whatever is decided is landed, not just written down.
+
+<!-- formalised by the Intake (Conception) Agent on 2026-08-28 from file:/tmp/claude-0/-home-user/fa910b96-e557-5368-a11e-13292ea18dec/scratchpad/intake_raw.md -->
diff --git a/tasks/point_solver_profiling_cells.md b/tasks/point_solver_profiling_cells.md
new file mode 100644
index 0000000..db4cc75
--- /dev/null
+++ b/tasks/point_solver_profiling_cells.md
@@ -0,0 +1,80 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autolens_profiling/point_solver_profiling_cells.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# PointSolver profiling cells: lensed quasar → cluster runtime tier → single/multi-source → multiplane
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+- autolens_workspace_test
+Themes:
+- point-source
+- profiling
+- cluster
+Difficulty: large
+Autonomy: supervised
+Priority: normal
+Status: formalised
+Consequence: glance
+Witness: Four cells exist in the repo taxonomy — a lensed-quasar runtime cell with fluxes, a cluster single-source runtime cell, a cluster multi-source multi-redshift runtime cell including a factor-graph fit, and the multiplane check promoted to the runtime tier — each with smoke early-exit, eager/JIT/vmap tiers and a pinned-likelihood drift record, and all four appear in the results/runtime dashboard.
+Review-minutes: 3
+Unattended: ready
+Epic: cluster-pointsolver-speed
+Origin-phase: cluster-strong-lensing phase 2
+Parent: draft/research/autolens_profiling/cluster_pointsolver_speed.md
+Filed: 2026-08-19 (backfilled from git)
+
+Transferred from Source & Cluster arc phase 2 on 2026-10-01. Owned by the
+PointSolver robustness/performance campaign; no blanket phase-1 completion gate. User request (verbatim): "Once satisfied extend to profiling examples in
+autolens_profiling but simple lensed quasar, cluster scale with single source, then
+cluster with multiple source redshifts and finally cluster with multiplane ray tracing."
+
+Survey findings (2026-08-19) — what exists vs what the four cells need:
+
+- Cluster multi-redshift/multiplane is ALREADY covered by
+  `scripts/cluster/likelihood_breakdown/{image_plane,source_plane}.py` (2 sources at
+  z=1.0/2.0, multi-plane). The real gap is the **`cluster/likelihood_runtime/` tier** —
+  cluster has breakdown/ and searches/ but no runtime tier, so cluster cells never appear
+  in the sweep-driver runtime dashboard.
+- **Lensed quasar**: no named cell; closest is `point_source/likelihood_runtime/
+  image_plane.py` (generic isothermal + 1 point source). No flux (`al.FitFluxes`) or
+  time-delay profiling cell exists at all — a quasar cell should include fluxes.
+- **Cluster single source**: `scripts/misc/simulators/cluster.py` is hardwired to 2
+  sources at 2 redshifts (source_redshifts=[1.0, 2.0] at :144-145) — needs a
+  single-source preset.
+- `point_source/` has no `likelihood_breakdown/` tier (imaging/interferometer/cluster
+  all have one).
+- `multi_dataset/` has no factor-graph/AnalysisFactor profiling cell — that is what a
+  multi-source cluster fit actually costs; add one in the cluster runtime tier.
+
+Follow the repo's taxonomy (`scripts/<dataset>/<task>/<model>.py`, canonical script shape
+per `point_source/likelihood_runtime/image_plane.py`: smoke early-exit, Timer,
+eager/JIT/vmap tiers, pinned-likelihood drift record). Trap on record: per-source
+`plane_redshift` MUST be passed to `solve` — it defaults to the tracer's final plane
+(#678 phase B); the cluster simulator's `jitted_solve_for(plane_redshift)` closure at
+cluster.py:339 is the template.
+
+Deliverable: 4 cells (quasar w/ fluxes; cluster single-source runtime; cluster
+multi-source multi-z runtime incl. factor-graph; multiplane already-covered check +
+runtime promotion), results in the standard results/runtime dashboard.
+
+<!-- formalised by the Intake (Conception) Agent on 2026-08-19 from file:/tmp/claude-1000/-home-jammy-Code-PyAutoLabs/483da28c-8c96-4c83-ad87-a43448ca2164/scratchpad/source_cluster_phases/phase02_point_solver_profiling_cells.md -->
+
+## Reconcile before issuing — approved 2026-10-01
+
+The August survey above is historical, not an assertion about current coverage.
+Reconcile point_source_image/ and point_source_source/ plus completed single-source
+CPU work before adding cells. Coordinate quasar/single-source gaps with that
+campaign; this prompt owns only missing coverage, never duplicate speed work.
+Cluster workloads and robustness/settings sweeps belong in autolens_profiling.
+Diagnostic timings may include explicitly failing cases; accepted speed and
+settings claims require independent accuracy/coverage, observable overflow and
+applicable JIT/vmap/gradient checks. Distil bounded numerical witnesses into
+autolens_workspace_test and assign required PR-smoke or scheduled/release CI
+coverage. The old four-cell list is an inventory to reconcile, not four issues
+to queue. No issue created by transfer.
diff --git a/tasks/point_source_image_plane_gpu_breakdown.md b/tasks/point_source_image_plane_gpu_breakdown.md
new file mode 100644
index 0000000..c189fab
--- /dev/null
+++ b/tasks/point_source_image_plane_gpu_breakdown.md
@@ -0,0 +1,212 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autolens_profiling/point_source_image_plane_gpu_breakdown.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Point-source A100 speed-up campaign: profile and optimize with the shared breakdown
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+- PyAutoArray
+- PyAutoLens
+Themes:
+- point-source
+- profiling
+- jax-gpu
+Difficulty: large
+Autonomy: supervised
+Priority: high
+Status: formalised
+Consequence: judge
+Review-minutes: 30
+Unattended: ready
+Filed: 2026-09-17
+Updated: 2026-09-30
+
+## Campaign contract (2026-09-19)
+
+This execution plan supersedes the earlier deliverable/gate wording retained below.
+This is one of two optimization campaigns in a three-task plan. The independent
+shared breakdown task is their common prerequisite.
+It is a phased campaign: at start-dev, issue only the next bounded phase (one task /
+one PR per member), retaining this prompt as the campaign intent until all phases
+are resolved. Do not attempt a cross-library, multi-PR campaign as a single task.
+Use start-dev and the applicable library/workspace worktree and ship procedures;
+obtain implementation-plan approval before source edits. No profiling or library
+implementation was performed during this consolidation.
+
+### Measurement and acceptance contract
+
+- Use @autolens_profiling for timing and versioned JSON + PNG evidence, with
+  README/dashboard regeneration. Correctness evidence belongs in library tests
+  and @autolens_workspace_test, not a timing-only assertion of scientific validity.
+- Record exact library/profiling commits, JAX/jaxlib versions, device, precision,
+  XLA flags, thread settings, model/data seed, source planes, solver grid/scale,
+  precision, neighborhood degree, capacity, warm-up, repetitions and cache state.
+  Historical numbers are leads, not comparable current baselines.
+- Separate tracing/lowering, compilation, first execution and warmed runtime.
+  Synchronize device outputs with block_until_ready; pass varying parameter
+  inputs through the production likelihood so constant folding cannot fake work.
+  Report repeated/interleaved A/B medians and dispersion on identical hardware,
+  both ms/likelihood and batch throughput, plus memory where relevant.
+- Retain a fused end-to-end production likelihood control. Prefix timing
+  differences can change fusion and contain noise: report residuals/negative
+  differences honestly and corroborate with a device trace rather than claiming
+  independently timed steps sum to the fused runtime.
+- Compare likelihood, image counts/positions, NaN padding/masks, magnification
+  filtering and source-redshift handling. Cover perturbed models, doubles/quads,
+  near-caustic/critical configurations and cluster multi-source/multiplane cases.
+  Preserve custom_jvp, eager/JIT/vmap parity and gradient correctness; distinguish
+  nondifferentiable image-topology transitions from failures in smooth regions.
+- Each iteration: baseline -> one hypothesis -> bounded prototype -> correctness
+  gate -> repeated full-likelihood A/B -> accept/reject -> reprofile and rank the
+  remaining bottleneck. State minimum detectable improvement from observed noise;
+  keep a change only for a repeatable material gain without correctness or
+  unacceptable compile/memory regressions. Record negative results too.
+- Stop when remaining cost is explained and no worthwhile measured lever remains,
+  or a concrete external blocker prevents the next experiment. Do not promise the
+  historical speedup, optimize indefinitely, or weaken correctness to hit a target.
+
+### Split request (verbatim)
+
+break into 3, with the first build the shared likelihood breakdown first, which I guess is CPU or GPU agnostic but maybe not, then PR them into main
+
+### Original consolidation request (verbatim)
+
+We did two reviews or assessments of the point source likelihood function recently one for CPU which was JAX and numba sparse (it concluded numba spaerse not worth it) and one for GPU. We may of made some prompts but I want you to assess all that the review put forward and ultimately end with two mind task or prompts, which could be epics, which will profile them with autolens_profiling and iteratively work on the speed up. One was focused in particular on writing an autolens_profiling likelihood_breakdown script, which it may of wrote or just planned, this would likely be the task before we go into specific CPU or GPU speed up
+
+## GPU campaign: depends on the shared breakdown, then measured A100 optimization
+
+### Evidence audit (2026-09-19)
+
+Fetched @autolens_profiling `origin/main` has the four point-source runtime
+scripts but no `scripts/point_source/likelihood_breakdown/` or corresponding
+point-source breakdown results. Its cluster breakdown times `solver.solve` as
+one block. The September 17 GPU prompt is a proposed investigation, not an A100
+bottleneck verdict. Existing historical A100 science/defaults evidence is not a
+substitute for this dedicated breakdown. No new A100 performance claim is made.
+
+The CPU sibling preserves reported scratch gains, not landed profiling artifacts.
+Its redundant vertex sort is distinct from necessary triangle deduplication;
+GPU benefit must be measured rather than extrapolated from CPU ratios.
+
+### Prerequisite — separate shared breakdown task
+
+Task 1 shipped in [autolens_profiling#293](https://github.com/PyAutoLabs/autolens_profiling/pull/293)
+and is recorded in `complete/2026/09/point-source-shared-breakdown.md`.
+Use its merged instrument, CPU reference results under
+`results/breakdown/point_source/` and reproducible baseline revisions; do not
+rebuild the harness here.
+The harness shares stages/schema across devices, but A100 execution and timing
+must be validated here. A CPU run is not proof of GPU correctness or performance.
+
+First validate the harness on A100 and acquire fp64 and separately labelled
+mixed-precision reference rows on the preserved unoptimized library commits.
+This remains possible if CPU improvements have already landed. Check numerical
+agreement against the CPU reference and production likelihood, recording
+precision-specific tolerances. Begin GPU optimization only after that baseline.
+The CPU campaign does not depend on this campaign finishing or obtaining a GPU slot.
+Coordinate changes to the shared solver with the CPU campaign to avoid conflicts.
+
+### Phase 1 — A100 bottleneck map
+
+Phase 0+1 (lean, preserved-revision reproduction skipped by human decision 2026-09-28) shipped in
+[autolens_profiling#353](https://github.com/PyAutoLabs/autolens_profiling/pull/353) and is recorded in
+`complete/2026/09/point-source-gpu-p01.md`. Phase 2 awaits the human go/no-go (memo in
+`results/notes/point_source_gpu_breakdown_2026_09.md`: one autolens_inference image-plane fit timing first).
+
+- Use the profiling repo's actual `hpc/README.md` and sync configuration. This
+  repo's `hpc/sync` has no push: update RAL code with git and library checkouts
+  through HPCPullPyAuto, verify revisions, submit through its supported driver,
+  and ingest pulled results from LOCAL_PULL_ROOT. Do not copy a generic
+  science-project push-submit recipe into this task.
+- Obtain a fused device timeline following the maintained trace precedent;
+  attribute kernels to source operations and distinguish launch/host overhead,
+  sorting/gathers, deflections, synchronization and device computation.
+- Test vmap batches 1/4/16 then larger only within measured memory headroom;
+  include the historical batch-3 control if useful. Use distinct parameter
+  inputs; report latency, throughput, peak memory and batch setup costs.
+- Separate fp64/mixed-precision, compilation/cold start/warm execution, simple
+  and representative cluster single-/multi-source/multiplane configurations.
+- Time primal likelihood and the custom_jvp gradient path separately, including
+  batched gradients. Confirm finite-difference agreement at smooth parameter
+  points before treating gradient throughput as usable.
+
+### Phase 2 — bounded optimization iterations
+
+1. Re-evaluate the CPU redundant-sort removal on A100, sharing the same library
+   fix if valid; coordinate ownership rather than implement a competing patch.
+2. If traces support launch-bound execution, prioritize vmap and benchmark the
+   batch break-even/memory limit. A faster per-item batch is not lower latency
+   for a serial caller; publish both and identify which caller can exploit it.
+3. If sorts/gathers dominate, target the measured operation and consider the
+   static-lattice precompute jointly with CPU; preserve containment and padding.
+4. If unrolled refinement drives compile cost, evaluate a JAX loop representation
+   only after checking step-shape changes and autodiff constraints. Compare
+   compile amortization and steady state; don't assume scan/loop is faster.
+5. If the implicit-gradient leg dominates, profile its Jacobian/linear algebra
+   before changing it; preserve the implicit derivative contract.
+6. If deflections dominate, route a bounded profile-specific library phase
+   informed by `scripts/lens/deflections/`; retain CPU regression measurements.
+
+For each lever publish accepted/rejected/deferred reasoning and re-profile after
+accepted changes. Initial launch-latency, sort dominance, unrolled-compile and
+Jacobian-cost claims remain hypotheses until phase 1 measures them.
+
+### Completion evidence
+
+Reuse the shared harness and its CPU reference; ship A100 baseline/final artifact pairs,
+trace-derived bottleneck map, batching/precision/compile/gradient tables, and
+`results/notes/point_source_gpu_breakdown_2026_09.md` with phase PRs and commands.
+Finish warranted optimization iterations, not just filing another list of
+follow-ups. Explicitly record rejected levers and residual bottlenecks. If A100
+access is unavailable, the separate shared task and CPU campaign can proceed
+but this campaign remains pending; never label CPU timing as A100 evidence.
+
+## Preserved September 17 assessment and provenance
+
+The following is historical context; the campaign contract above governs new work.
+
+# Point-source image-plane chi-squared on the A100: likelihood breakdown, bottleneck map, speed-up levers
+
+User request (verbatim, 2026-09-17): "In autolens_profiling, we have done lots of work
+speeding up imaging and interferometer on CCD. Now, I want us to speed up the point_source
+image plane chi-squared, with this issue focusing on JAX GPU Speed up, mostly using the A100s
+on RAL to perform the profiling. Follow the same strurture as imaging and interferomerter,
+produce the likelihood breakdown (which likely will ned to be made for the steps of the point
+solver? I'm not sure if we ever made one) and focus on bottlenecks and so on."
+
+Survey (2026-09-17):
+
+- `point_source/` has a `likelihood_runtime/` tier (image_plane, image_plane_solved,
+  source_plane, source_plane_solved) but **no `likelihood_breakdown/` tier** and **no A100
+  row** — the only numbers are local CPU (image_plane_solved: 39 ms/JIT call, 28 ms under
+  vmap, v2026.7.23.1). `cluster/likelihood_breakdown/image_plane.py` is the closest
+  precedent but times `solver.solve` as one block; nobody has ever opened the solver loop.
+- The JAX solve is `AbstractSolver.steps()`: `n_steps = ceil(log2(scale/precision))` (=8 for
+  the 0.2"/0.001" cell) unrolled iterations of ray-trace → `containing_indices` (padded to
+  `MAX_CONTAINING_SIZE=15`) → `for_indexes` → `neighborhood()`^degree → `up_sample()`, each
+  of the last three calling `remove_duplicates` = `jnp.sort` + `jnp.unique(size=…)`, then
+  `_filter_low_magnification`, then the all-pairs permutation log-sum-exp chi-squared,
+  wrapped in the padded `custom_jvp` solve (`implicit_diff.py`).
+- Hypotheses to test on the A100: (a) the call is launch-latency / host-sync bound (tens of
+  tiny sort/unique/gather kernels per step, ~8 steps) rather than FLOP bound, so vmap over
+  likelihood evaluations is the first-order lever; (b) sort-based `unique` dominates device
+  time; (c) 8× unrolled steps inflate compile time; (d) the implicit-diff gradient leg
+  (vmap'd Jacobian) is a separate cost centre for gradient searches.
+
+Deliverable: a `point_source/likelihood_breakdown/image_plane.py` cell that opens the solver
+loop (per-step rows + chi-squared row, prefix-walk by successive differences per the imaging
+harness), A100 fp64 + mp rows and a local-CPU row in `results/breakdown/point_source/`, a
+fused-program device-timeline trace leg (fixed_light_trace.py precedent) attributing GPU
+kernels to library source lines, vmap-batch scaling on the A100, a compile-time row, a
+gradient leg, README dashboard regen, and a `results/notes/point_source_gpu_breakdown_2026_09.md`
+verdict ranking the levers with measured evidence. Library levers (PyAutoArray triangles /
+PyAutoLens solver) are filed from the verdict as follow-up prompts, library-first.
+
+Related: `draft/research/autolens_profiling/point_solver_profiling_cells.md` (cluster arc
+phase 2 — more point-source cells, not a GPU campaign; do not merge scopes).
diff --git a/tasks/point_source_source_plane_chi_squared_speed.md b/tasks/point_source_source_plane_chi_squared_speed.md
new file mode 100644
index 0000000..4140772
--- /dev/null
+++ b/tasks/point_source_source_plane_chi_squared_speed.md
@@ -0,0 +1,291 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autolens_profiling/point_source_source_plane_chi_squared_speed.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Point-source source-plane chi-squared speed-up campaign — remaining candidates (phases 1–2e shipped)
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+- PyAutoLens
+- PyAutoFit
+Themes:
+- point-source
+- profiling
+- jax
+Difficulty: large
+Autonomy: supervised
+Priority: low
+Status: formalised
+Consequence: judge
+Review-minutes: 25
+Unattended: needs-slicing
+Epic: point-source-cpu-speed
+Lane: any
+Filed: 2026-09-26
+Updated: 2026-09-28
+Parent-record: complete/2026/09/point-source-source-plane-p2e.md
+
+## Runtime refresh + A100 vmap row shipped — merged c483417 (2026-09-28)
+
+- **Slice:** autolens_profiling#349 → PR #351, merged c483417 (2026-09-28); record
+  `complete/2026/09/source-plane-runtime-refresh.md`. PyAutoFit#1649 / PyAutoLens#752 are released
+  in 2026.9.27.2 (wiki corrected).
+- **Rows (euclid-ral-gpu-2, qualified):** RAL CPU job 366911 single JIT 0.258 ms; A100 job 366912
+  vmap(b64) 5.6 µs/call ≈ 114× the single call; launch-bound through b1024 (0.29 µs/call, diag 366914).
+  The A100 cell `single_jit` (0.642 ms) is a warm-up artefact (steady 0.267 ms) → filed
+  `draft/bug/autolens_profiling/runtime_cell_single_jit_gpu_warmup.md`. Source-plane A100 leg added to
+  the release sweep.
+- **Remaining candidate:** blackjax NUTS / SMC forward-mode `value_and_grad` — stays parked until
+  `lens/autolens_inference/scripts/point_source/searches/` has a point-source search leaf (its
+  admission bar is the likelihood's share of a fit + eval count).
+
+## Phases 1–2e shipped — campaign core complete; remaining candidates parked (2026-09-27)
+
+- **Phase 2e merged:** autolens_profiling#336 at `17e2596` (issue #334). Record
+  `complete/2026/09/point-source-source-plane-p2e.md`. The real `af.MultiStartAdam` step runs forward
+  mode at 0.36–0.58× reverse on the quiet RAL EPYC (matching phase 2c); L24 compile 95.9 → 11.2 s.
+- **State of the likelihood:** the forward call is at its CPU dispatch floor (phase 2a); the gradient
+  takes the cheaper AD mode by default once PyAutoFit#1649 / PyAutoLens#752 are released.
+- **Remaining candidates (not started; issue one at a time only if wanted):**
+  1. blackjax NUTS / SMC forward-mode `value_and_grad` (they differentiate the log-density
+     themselves; phase 2d scoped them out).
+  2. A100 `vmap` throughput row (the single call is launch-bound; batch throughput is the GPU number
+     that matters).
+- **Filed alongside:** `draft/bug/autoarray/galaxy_duplicate_pytree_registration_after_jax_fitness.md`,
+  `draft/bug/autogalaxy/power_law_multipole_m1_singular_at_isothermal_slope.md`.
+
+## Phase 2d shipped (library) — next is phase 2e (2026-09-27)
+
+- **Merged:** PyAutoFit#1649 at `867af1c` then PyAutoLens#752 at `b3c9b68` (issue PyAutoFit#1648);
+  both pending release. Record `complete/2026/09/point-source-gradient-mode.md`. `af.Analysis.gradient_mode`
+  defaults to `"reverse"`; `AnalysisPoint` declares `"forward"`; `af.MultiStartAdam(gradient_mode=...)`
+  overrides; helpers in `autofit/jax/gradient.py`.
+- **Next: phase 2e (workspace, the re-filed remainder of phase 2d's plan)** — autolens_profiling
+  `scripts/point_source_source/likelihood_breakdown/gradient_mode_library_ab.py`: the real
+  `MultiStartGradient` step on the phase-2c L5 and L24 rungs with `gradient_mode="forward"` vs
+  `"reverse"`, RAL gpu-node CPU + A100, confirming the phase-2b/2c gain arrives through the library
+  path; campaign-note section. Run against the library mains now (merged) or after the release.
+- **Later:** blackjax NUTS / SMC forward-mode `value_and_grad` (phase 2d scoped it out).
+- **Carried:** intake — `Galaxy` duplicate PyTreeDef registration (JAX `Fitness` then
+  `register_tracer_classes` in one process); `PowerLawMultipole` m=1 singular at slope 2. RAL cleanup
+  (`/mnt/ral/jnightin/p2d_check`, p2a/p2b/p2c worktrees).
+
+## Phase 2c shipped — next is phase 2d (2026-09-27)
+
+- **Merged:** autolens_profiling#331 at `4c267b7` (issue #329). Record
+  `complete/2026/09/point-source-source-plane-p2c.md`; cell
+  `scripts/point_source_source/likelihood_breakdown/gradient_mode_crossover.py`; design memo in the
+  campaign note ("Phase 2c" → "Design memo").
+- **Result:** no forward/reverse crossover through n = 24 (solved) / 27 (plain) on any host or call
+  shape; forward mode's single-call lead grows with model size (0.41 → 0.22 on the quiet RAL EPYC,
+  0.64 → 0.36 on A100) and rev compile grows 3.4 → 80 s vs fwd ≤ 11 s. Cause: reverse-over-forward
+  through the inner jacfwd lensing Hessian. An `n_params` threshold is the wrong design.
+- **Next: phase 2d (human decision 2026-09-27) = analysis-declared `gradient_mode`, library-first
+  (PyAutoFit, then PyAutoLens):** `af.Analysis.gradient_mode = "reverse"` (default, no behaviour
+  change elsewhere); `AnalysisPoint` declares `"forward"`; a search keyword overrides. One PyAutoFit
+  helper returns `value_and_grad` (reverse) or `jacfwd(has_aux=True)` over the flat parameter vector
+  (forward) and feeds `Fitness.grad` (`autofit/non_linear/fitness.py:933`) and the multi-start
+  gradient search (`autofit/non_linear/search/mle/multi_start_gradient/search.py:974`, `:1072`,
+  vmapped `:1089`). blackjax NUTS/SMC are a later step. Use the flat vector, not the `ModelInstance`
+  pytree (linked priors give more leaves than parameters). Gates: fwd ≡ rev parity tests on a
+  point-source and a non-point-source analysis, GPU regression check, then refresh the profiling rows.
+- **Carried:** intake PyAutoGalaxy bugs — `jax.grad` NaN at exactly (0,0) for ExternalShear /
+  multipole comps / ell_comps; `PowerLawMultipole` m=1 singular at slope 2; `Isothermal.convergence_2d_from`
+  (and `shear_yx_2d_from`) not jit-traceable with traced ell_comps (fixed: PyAutoGalaxy#633, record
+  `complete/2026/09/isothermal-convergence-jit.md`). RAL p2a/p2b/p2c worktrees + bundles to remove.
+
+## Phase 2b shipped — next is phase 2c (2026-09-27)
+
+- **Merged:** autolens_profiling#327 at `4dc05a4` (issue #325). Record
+  `complete/2026/09/point-source-source-plane-p2b.md`; cell
+  `scripts/point_source_source/likelihood_breakdown/backward_pass_ab.py` (the A/B harness to extend).
+- **Verdict:** forward-mode gradient (`jax.jacfwd` over the flat 5-vector) is **GO** — −46 % solved /
+  −38 % plain on the quiet RAL gpu-node EPYC 7702 CPUs (human re-based the rule off the busy 8490H),
+  −24–34 % on A100, A100 compile 4.1 → 1.8 s. Analytic SIE Hessian adds nothing on top of `fwd`;
+  jacrev Hessian ordering is a no-op.
+- **Next: phase 2c (human decision 2026-09-27) = crossover study, workspace-only.** Extend the A/B
+  harness from 5 to ~20 free parameters (external shear, multipoles, a second lens; single-source
+  only) and measure `fwd` vs `rev` per `n_params` on RAL CPU + A100 to find where forward mode stops
+  winning. Only then design the PyAutoFit gradient-entry-point switch (conditional on `n_params`)
+  from the measured crossover — that design is a human decision before any library phase.
+- **Carried:** `Isothermal.convergence_2d_from` / `shear_yx_2d_from` not jit-traceable with traced
+  `ell_comps` (both fixed by PyAutoGalaxy#633 — `shear_yx_2d_from` routes through `convergence_2d_from`;
+  jit == NumPy verified 2026-09-27, pre-#633 source raises TracerArrayConversionError); RAL p2a/p2b
+  worktrees + bundle to remove.
+
+## Phase 2a shipped — remainder re-filed (2026-09-27)
+
+- **Merged:** autolens_profiling#323 at `9c0203e` (issue #322). Record
+  `complete/2026/09/point-source-source-plane-p2a.md` (it folds the previous version of this
+  prompt, including the phase 1 section below).
+- **Measured (RAL CPU 8490H, quiet — the reference):** fused solved 0.1465 ms / plain 0.1450 ms;
+  `value_and_grad` 2.26× forward (+0.185 ms), 4× compile. Pytree / flat_vector = 1.361
+  [1.343, 1.385] → 0.0385 ms saved (plain 0.0452); flat_leaves ≈ flat_vector, so the cost is the
+  Python flatten of `ModelInstance`. A100 launch-bound 0.222 / 0.265 ms.
+- **Human decision (2026-09-26):** NO-GO on the PyAutoFit flatten fast path (step (b) below) —
+  relative 26–31 % passes, absolute 0.05 ms bar fails. Steps (a), (b) and (d) below are done.
+- **Next: phase 2b = the backward-pass lever** (step (c)): jacfwd/jacrev ordering and an analytic
+  SIE Hessian study on the solved likelihood; library-first (PyAutoLens) only if a lever appears.
+- **Carried:** laptop rows never refreshed (RAL is the reference); `scripts/misc/wall/check_submits.py`
+  misses `python3 -u`; `hpc/sync pull` skips batch_cpu logs; RAL worktree
+  `/mnt/ral/jnightin/autolens_profiling_wt/point-source-source-plane-p2a` to remove.
+
+## Phase 1 shipped — remainder re-filed (2026-09-26)
+
+- **Merged:** autolens_profiling#317 at `f79ebf2` (after the folder split #318, `a5e3cdd`, which
+  moved `scripts/point_source/` → `scripts/point_source_image/` + `scripts/point_source_source/`;
+  paths below that still say `scripts/point_source/` now live under `scripts/point_source_source/`).
+  Record `complete/2026/09/point-source-source-plane-breakdown.md`; campaign note
+  `results/notes/point_source_source_plane_campaign.md` (ranked residue + "carried to cluster epic").
+- **Measured headline** (laptop CPU fp64, 8 threads, median of 500): fused solved 0.438 ms /
+  plain 0.457 ms; `value_and_grad` 2.43× forward (4× compile); dispatch floor 0.07–0.09 ms with
+  the pytree argument (0.019 ms scalar); XLA merges the repeated Hessian/precision evaluations
+  (hypothesis #1 answered: CSE does merge them — tidiness only). The plain-path
+  `TracerArrayConversionError` guard was retired (plain fit JITs end-to-end). **Laptop host load
+  was ~16 on 8 cores — re-take on an idle host or RAL before ranking.**
+- **This prompt now carries phases 2+ only.** Next steps, in order:
+  1. **(a)** idle-host or RAL re-run of the `point_source_source/source_plane` breakdown row under a
+     new label (plus the unrun `hpc_ral_cpu_fp64` / `hpc_a100_fp64` rows; commands in the note).
+  2. **(b) Phase 2 = the top residue lever — fixed per-call cost / `ModelInstance` pytree
+     flattening** (~0.1 ms of a ~0.3–0.4 ms call; flat-leaf or single-vector input cut the floor to
+     0.15–0.20 ms in a scratch probe). Likely PyAutoFit-scoped (`register_model` pytree
+     flattening) — library-first.
+  3. **(c)** backward-pass cost (2.43× forward, 4× compile) as the phase 3 candidate.
+  4. **(d)** A100 launch-bound / vmap-throughput row — unmeasured.
+
+## User request (verbatim, 2026-09-26)
+
+Continue on going work to speed up the JAX source plane chi squared point solver, which we have been working on recently. should be an epic laid out for it I think? We have work on the image plane one, but maybe not the source plane chi squarded, so have a look at what is available. first task will be to write a likelihood_breakdown and then speed up from there, autolens_profiling gives a clear overview of the whole process and task and steps and design which you can use from other examples like imaging and the image plane point source.
+
+## Scope steer (human, 2026-09-26): single-source only
+
+This campaign is **single-source only** — the `scripts/point_source/` use case. No two-source /
+cluster setups, rows, submits or levers; `scripts/cluster/` is never touched. The cluster use
+case is epic `cluster-pointsolver-speed` (its own data and `scripts/cluster/likelihood_breakdown/`
+baseline). Cluster-specific evidence found on the way is written into a short "carried to
+cluster epic" note in the hand-off, never acted on. Epic tag is `point-source-cpu-speed`
+(`cluster-strong-lensing` is the unrelated Source & Cluster arc).
+
+## Survey (2026-09-26): what exists
+
+- The **image-plane** PointSolver campaign is fully laid out and three phases in:
+  shared breakdown instrument `scripts/point_source/likelihood_breakdown/image_plane.py`
+  (record `complete/2026/09/point-source-shared-breakdown.md`, note
+  `results/notes/point_source_shared_likelihood_breakdown.md`), phases 1-3 shipped
+  (`complete/2026/09/point-source-cpu-p{1,2,3}.md`, ledger
+  `results/notes/point_source_cpu_campaign.md`), phase 4 drafted at
+  `complete/2026/09/point-source-cpu-p4.md`.
+- The **source-plane** chi-squared (`al.FitPositionsSource` / `al.FitPositionsSourceSolved`,
+  Lenstool's default likelihood, no lens-equation solve) has **no campaign, no epic entry and
+  no breakdown instrument**. What exists:
+  - runtime cells `scripts/point_source/likelihood_runtime/source_plane.py` (plain; its
+    end-to-end JIT is guarded by a `try/except TracerArrayConversionError` fallback to a
+    ray-trace-only prefix) and `source_plane_solved.py` (full pipeline JITs;
+    `results/runtime/point_source/source_plane_solved/…v2026.7.23.1.json`: eager 10.9 ms,
+    single-JIT 0.34 ms/call, vmap(3) 0.12 ms/call, LL 0.5986504555530896);
+  - `scripts/cluster/likelihood_breakdown/source_plane.py` (593 lines, v2026.7.23.1): a
+    closure-based per-step decomposition of the 13-component (2 dPIE + 10 scaling-tier dPIE +
+    1 NFW), two-source (z=1,2) multi-plane cluster case — steps: multi-plane ray trace of the
+    observed positions, Hessian magnification at every observed position
+    (`magnification_2d_via_hessian_from`, re-evaluates the full deflection stack several
+    more times per position), magnification-weighted chi-squared, log-likelihood assembly;
+    plus a plain-vs-solved eager per-system comparison (steps 5-6). README row
+    `cluster/source_plane` local_cpu_fp64 step-sum 4.5 ms. It closes over a fixed tracer
+    (params are compile-time constants, so constant folding can fake work), is CPU-only in
+    practice, and its docstring records the plain end-to-end fit as JIT-blocked.
+- Issue #657 series (`project` memory: point-source solved likelihoods): `PointSolved` +
+  `FitPositionsSourceSolved` regression literal SourceSolved −94.70750993 (cluster,
+  vmap==eager exact); the paper is Lombardi 2024 arXiv:2406.15280 §5.1.
+- The point-source defaults campaign (`results/notes/point_source_defaults_campaign.md`,
+  PyAutoLens#678) adopted tensor source-plane weighting + solved centres as defaults, and
+  found gradient searches not yet competitive at cluster scale on the source-plane objective —
+  so per-call cost AND gradient cost both matter here.
+
+## Campaign contract
+
+This is a phased campaign, the source-plane sibling of the image-plane CPU campaign. At
+start-dev issue ONLY the next bounded phase (one task / one PR per member), retaining this
+prompt as the campaign intent until every phase is resolved. The image-plane campaign's
+**measurement and acceptance contract** (in
+`complete/2026/09/point-source-cpu-p4.md`, "Campaign contract")
+governs verbatim: record commits / JAX versions / device / precision / threads / seeds;
+separate lowering, compile, first call and warmed runtime; block_until_ready; vary
+parameters through the production likelihood so constant folding cannot fake work; retain a
+fused end-to-end production-likelihood control; report interleaved A/B medians + dispersion
+on identical hardware; cover perturbed models, doubles/quads and near-caustic cases; preserve custom_jvp / eager-JIT-vmap parity / gradient
+correctness; each iteration = baseline → one hypothesis → bounded prototype → correctness
+gate → repeated A/B → accept/reject → reprofile; record negative results; stop when the
+residue is explained.
+
+### Phase 1 — SHIPPED (see above)
+
+_Original phase-1 heading: shared CPU/GPU likelihood breakdown instrument._
+
+Build `scripts/point_source/likelihood_breakdown/source_plane.py` (single-source `simple` instrument only), structurally mirroring
+`image_plane.py` (same result-JSON contract, `--config-name` hardware rows, provenance
+block, JIT-phase split, numerical controls, PNG, README dashboard regeneration via
+`build_readme.py`), for the production `AnalysisPoint.log_likelihood_function` with:
+
+1. **Primary path** `al.ps.PointSolved` + `al.FitPositionsSourceSolved` (the adopted
+   default); **separately labelled control** free-centre `al.ps.PointFlux`/`al.ps.Point` +
+   `al.FitPositionsSource`. Never conflate the two variants.
+2. **One instrument:** the seeded four-image `simple` dataset (as `image_plane.py`). At this scale
+   the call is ~0.3 ms so fusion/launch overhead may dominate — report it honestly.
+3. **Cumulative prefix boundaries through the fused production likelihood** (not a fixed
+   closure): (a) multi-plane ray trace of the observed positions to each source plane;
+   (b) Hessian magnification at each observed position; (c) β* solve (solved path) / model
+   centre lookup (plain path); (d) magnification-weighted chi-squared + noise normalisation;
+   final row = fused likelihood minus last prefix. Parameters enter as a registered
+   `af.ModelInstance` pytree and are varied per call. Retain the telescoping-sum caveat
+   from the image-plane note (prefix rows are not independently additive).
+4. **Controls:** eager ≡ JIT ≡ vmap(2) log-likelihood parity, hard-coded regression
+   literals (refresh procedure documented), full-model `jax.grad` finite AND non-zero
+   (`autofit.jax.register_model` — an unregistered model silently yields all-zero grads),
+   and a **gradient-cost row** (ms per `value_and_grad` call vs forward) because the
+   defaults campaign found gradient searches uncompetitive at cluster scale.
+5. **Plain-path JIT status:** ground on the live stack whether `FitPositionsSource`
+   end-to-end still raises `TracerArrayConversionError` (the runtime cell's guard and the
+   cluster breakdown's docstring both say blocked; PyAutoArray#414 claimed the `xp` fix).
+   If it JITs, drop the fallback from the runtime cell in the same PR; if not, record the
+   exact failing frame as the first ranked lever for phase 2.
+6. **Results:** `results/breakdown/point_source/source_plane_{local_cpu_fp64,…}.json/.png`
+   for both cells on this laptop (record `NPROC`, XLA threads) and, if RAL is reachable,
+   `hpc_ral_cpu_fp64` + `hpc_ral_a100_fp64` rows under `--nodelist`-pinned submits with the
+   CPU model recorded; a campaign note
+   `results/notes/point_source_source_plane_campaign.md` with the ranked residue for phase 2.
+   Smoke: honour `AUTOLENS_PROFILING_SMOKE=1`; regenerate READMEs (`build_readme.py --check`).
+7. **Out of scope for phase 1:** any library edit beyond the plain-path fallback removal;
+   no speed-up levers.
+
+### Phases 2+ (to be ranked by the phase-1 breakdown, issued one at a time)
+
+Candidate levers, unmeasured — the phase-1 residue ranks them:
+- **Hessian magnification / precision-tensor re-evaluation** — the plain path evaluates
+  `magnifications_at_positions` twice per likelihood and the solved path rebuilds the jacfwd
+  precision tensor three times (plus `_beta_hat` twice); whether XLA CSE merges them is hypothesis #1; a `jax.jacfwd` of the single ray trace, or
+  reusing the ray-trace deflections, may remove those evaluations (bit-identical-or-tolerance
+  gate against the finite-difference Hessian, near-critical positions included).
+- **Plain-path JIT block** (`Grid2DIrregular` `xp` propagation) if still present.
+- **Gradient cost** — if `value_and_grad` ≫ forward, profile the backward pass
+  (jacfwd-of-deflections Hessian-of-Hessian).
+- **vmap batch throughput** on A100 (launch-bound at 0.3 ms/call).
+Each phase: library-first (PyAutoLens/PyAutoArray) then refresh the profiling rows under a
+new label; GPU regression check on every shared library change.
+
+## Where the code lives
+
+- `autolens_profiling/scripts/point_source/` — cells (`scripts/cluster/` is out of scope: epic `cluster-pointsolver-speed`).
+- `PyAutoLens/autolens/point/fit/` — `FitPositionsSource`, `solved.py` (`SolvedCentre`), `autolens/lens/` — multi-plane tracer, `LensCalc.magnification_2d_via_hessian_from`.
+- `PyAutoGalaxy/autogalaxy/operate/lens_calc.py` — `hessian_from` / `magnification_2d_via_hessian_from` (jacfwd path).
+
+## Related
+
+- `complete/2026/09/point-source-cpu-p4.md` — image-plane sibling campaign (contract source), same epic.
+- `draft/research/autolens_profiling/cluster_pointsolver_speed.md` — the cluster epic that receives any carried cluster evidence.
+- `draft/research/autolens_profiling/point_solver_profiling_cells.md`, `point_source_image_plane_gpu_breakdown.md` — related point-source prompts.
+
diff --git a/tasks/pointsolver_cpu_speed_campaign_remainder.md b/tasks/pointsolver_cpu_speed_campaign_remainder.md
new file mode 100644
index 0000000..7631d7f
--- /dev/null
+++ b/tasks/pointsolver_cpu_speed_campaign_remainder.md
@@ -0,0 +1,67 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autolens_profiling/pointsolver_cpu_speed_campaign_remainder.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Point-source (single-source) CPU campaign — carried leftovers and completion evidence
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+Themes:
+- point-source
+- profiling
+Difficulty: small
+Autonomy: supervised
+Priority: low
+Status: formalised
+Epic: point-source-cpu-speed
+Filed: 2026-09-27
+Parent-record: complete/2026/09/point-source-cpu-p4.md
+
+## Status
+
+Split out of `point-source-cpu-p4` at close-out on 2026-09-27. Phase 4a shipped in
+[autolens_profiling#321](https://github.com/PyAutoLabs/autolens_profiling/pull/321), record
+`complete/2026/09/point-source-cpu-p4.md`, whose `## Original prompt` holds the full campaign contract.
+The phase-4 code levers have their own prompts:
+- phase 4b: shipped 2026-09-27, record `complete/2026/09/pointsolver-step0-gather.md` (PyAutoArray#580, autolens_profiling#330)
+- phase 4c: shipped 2026-09-27, record `complete/2026/09/pointsolver-mcs-headroom.md` (PyAutoArray#584, PyAutoLens#753, autolens_profiling#335)
+- the extent warning: `complete/2026/10/pointsolver-extent-sanity-check.md`
+- per-package extents: `draft/feature/autolens_workspace/pointsolver_grid_extent_per_package.md`
+
+This prompt holds only the phase 1–3 leftovers that no other prompt carries. Pick it up after the
+phase-4 levers resolve.
+
+## Carried leftovers
+
+- **RAL cleanup** (once the release is synced):
+  - the `/mnt/ral/jnightin/autolens_profiling_wt/{PyAutoArray,PyAutoLens}_point-source-cpu-p{2,3}` clones
+  - the `point-source-cpu-p3` and `point-source-cpu-p4` RAL worktrees
+  - `_p2_untracked_backup_20260924`
+- **Test placement:** the PyAutoLens JAX unit test `test_autolens/point/triangles/test_static_lattice_jax.py`
+  (~85–100 s) departs from the no-JAX-in-unit-tests convention. Consider moving it to
+  autolens_workspace_test. Phase 4b pins its tie test, so do this only after 4b ships.
+- **Library candidate (PyAutoFit / PyAutoLens):** `jax.grad` of an `AnalysisPoint` likelihood is
+  silently all-zero without `autofit.jax.register_model(model)`. The fix is to raise or warn. File it
+  via intake if no bug prompt covers it by then.
+- **CI smoke coverage** for the breakdown cells: `vertex_dedup_ab.py`, `static_lattice_ab.py`,
+  `solver_config_sweep.py`.
+- **Unmeasured controls:**
+  - the `--xla_disable_hlo_passes=constant_folding` A/B on the post-4b code
+  - repeated (median) compile timings, if compile becomes a gate. Phase 3 had one cold compile per route.
+- **Phase 4b leftovers** (record `complete/2026/09/pointsolver-step0-gather.md`):
+  - `nopad` is dead weight in PyAutoArray `_STEP0_CONTAINMENT`. Delete it in a later library cleanup; `structured` is the default.
+  - The RAL scratch copies `/mnt/ral/jnightin/autolens_profiling_wt/pointsolver-step0-gather` and
+    `/mnt/ral/jnightin/autolens_profiling_wt/PyAutoArray_pointsolver-step0-gather` were left in place. Remove them with the RAL cleanup above.
+  - The quotable 8490H row (job 357321) ran on a loaded node (load ~200). Re-run it on a quiet node if absolute ms against phase 4a are ever wanted; the ratios stand.
+- **Phase 4c leftovers** (record `complete/2026/09/pointsolver-mcs-headroom.md`):
+  - The A100 vmap-4 cell (job 359102) read +8 % on a single cell. This is likely noise; re-run it before quoting it.
+  - The draw-12 fold-line attribution was carried from 4a and not re-verified in 4c.
+  - The RAL scratch worktree `/mnt/ral/jnightin/autolens_profiling_wt/pointsolver-mcs-headroom` was left in place. Remove it with the RAL cleanup above.
+- **Campaign completion evidence** (campaign contract): baseline/final comparison, every candidate
+  disposition, and a GPU regression check for any shared library change. Record them in
+  `results/notes/point_source_cpu_campaign.md` once the phase-4 levers are resolved, then close the epic.
diff --git a/tasks/post_certified_solver_likelihood_breakdown.md b/tasks/post_certified_solver_likelihood_breakdown.md
new file mode 100644
index 0000000..b75dbf2
--- /dev/null
+++ b/tasks/post_certified_solver_likelihood_breakdown.md
@@ -0,0 +1,58 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autolens_profiling/post_certified_solver_likelihood_breakdown.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Assess remaining likelihood bottlenecks after certified solver integration on CPU and GPUs
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+Difficulty: large
+Autonomy: supervised
+Priority: normal
+Status: formalised
+Consequence: judge
+Witness: All nine backend/mesh combinations have measured whole-call and exclusive breakdown records that reconcile within 5% or quantify instrumentation uncertainty; a ranked optimization assessment states measured contribution, correctness constraints and whole-call speedup ceiling for each candidate.
+Review-minutes: 20
+Unattended: needs-slicing
+Blocked-by: PyAutoMind/complete/2026/09/certified-positive-solver.md — PyAutoArray#567 merged 2026-09-23, unreleased — blocked on the release that ships it (phase B policy shipped in complete/2026/09/certified-solver-phase-b.md; production default flip is draft/feature/autoarray/certified_solver_scalar_default_flip.md) — RESOLVED 2026-09-27: PyAutoArray#567 released 2026.9.26.1; the default-flip draft was retired unbuilt (`complete/2026/09/certified-solver-scalar-default-flip.md`) and the library default stays `pdip`, so "post-certified" means the opt-in solver as configured, not a new default
+Superseded-in-part: GPU (A100 / RTX) columns measured by autolens_profiling#268 (hst-gpu-non-solver-residue phase 1, 2026-09-16); numba CPU columns are the fixed-lens-light-numba-cpu campaign's (#263/#265/#267)
+
+## Original request
+
+Now intake a follow up issue which is that once the ceritifed solver is implemented, to look at the likelihood⠂breakdown and work out if optimization of the other⠐parts which are now the bottlenecks is
+  possible. This should be⠄done for numba CPU⢀and JAX GPU (A100 and RTX), for rectangular, delaunay and delaunay NN.
+
+## Objective and dependency
+
+After the production certified solver and its dispatch are implemented and validated in the installed stack, use @autolens_profiling to measure where the likelihood spends time and assess feasible optimization of the remaining bottlenecks. This is one bounded assessment task; library implementation is a separate follow-up. Do not substitute the earlier solver monkeypatch or assume that certified solving is selected or fastest on native CPU.
+
+Dependency: PyAutoMind/complete/2026/09/certified-positive-solver.md (PyAutoArray#567, merged 2026-09-23, released 2026.9.26.1). Record the merged library revision and actual production-selected solver/configuration for every measurement.
+
+## Required measurements
+
+Cover all nine combinations: Numba CPU, JAX GPU A100, JAX GPU RTX, each with rectangular, Delaunay and DelaunayNN. Confirm each path is available; missing hardware/path is an explicit unfinished measurement gap, never an extrapolated result.
+
+Use matched fixed-lens-light source-only fits, one controlled representative dataset/configuration, fp64 reference, and roughly 1500 source pixels (record actual built and solved counts, particularly rectangular edge removal). Record CPU/BLAS/Numba threads, GPU model, preprocessing provenance and library/configuration revisions. Separate setup/compilation from warmed steady-state execution and report timing variability.
+
+Measure whole-likelihood time and a same-process exclusive decomposition, including the solver so the proposed bottleneck shift is verified. Attribute mesh/mapper construction, ray tracing, PSF/blurring, mapping/data-vector and curvature/regularization assembly, determinants/evidence, and repeated builds, copies, host/device transfers or callbacks where observed. Avoid overlapping terms; reconcile to whole-call time within 5% or quantify unexplained time, synchronization and instrumentation overhead. Kernel times subtracted across different runs are not a definitive breakdown; check that instrumentation has not materially changed GPU fusion/execution.
+
+Reuse the ongoing CPU work's dense/sparse preprocessing provenance and fixed-light subtraction re-bake/parity checks where applicable. Preserve positivity, edge constraints, reconstruction/evidence and backend parity. Fixing light uses precomputed solved intensities; state its preparation cost separately.
+
+## Deliverable and boundaries
+
+Publish machine-readable measurements and a concise per-hardware/per-mesh verdict ranking opportunities by measured contribution, expected whole-call speedup ceiling, implementation cost and correctness tradeoffs. Distinguish demonstrated improvements from plausible hypotheses. Bounded profiling-harness experiments may test feasibility; any proposed source-library change becomes a separately scoped implementation task. Wider source-size, dataset, precision or batching sweeps require evidence-driven follow-ups. No speedup is promised.
+
+Reconcile overlap at planning time with these existing Mind records, reusing useful results/instrumentation without duplicating or silently replacing ongoing work:
+
+- complete/2026/09/hst-gpu-non-solver-residue.md (COMPLETE 2026-09-27): broader HST/A100 optimization campaign, predating production integration.
+- complete/2026/09/fixed-lens-light-numba-cpu.md: completed CPU campaign; HST Delaunay evidence is reusable, not full nine-cell coverage.
+- complete/2026/09/fixed-light-numba-phase1.md (harness shipped, legs unrun) and complete/2026/09/fixed-light-numba-solver.md: the CPU measurement/decomposition.
+
+Use autolens_profiling/results/notes/fixed_lens_light_{library_path,hardware,source_pixel_scaling,verdict}_2026_09.md as historical context, not fresh production baselines. Formalizes the non-solver bottleneck seed from autolens_profiling#259 in PyAutoMind/ideas.md.
+
+<!-- formalised by the Intake (Conception) Agent on 2026-09-15 from file:tmp/post_certified_likelihood_breakdown_intake.md -->
diff --git a/tasks/profile_lens_aggregator_needs_config_dir.md b/tasks/profile_lens_aggregator_needs_config_dir.md
new file mode 100644
index 0000000..5b2b25a
--- /dev/null
+++ b/tasks/profile_lens_aggregator_needs_config_dir.md
@@ -0,0 +1,34 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/bug/workspaces/profile_lens_aggregator_needs_config_dir.md
+Migrated-at: 2026-10-03
+
+Original task retained verbatim below. Current scheduling state: `campaigns.yaml`. Historical paths: `migration.yaml`.
+
+---
+
+# profile_lens_aggregator.py cannot run from the autolens_workspace_developer root: no config/ directory
+
+Type: bug
+Target: autolens_workspace_developer
+Repos:
+- autolens_workspace_developer
+Difficulty: small
+Autonomy: safe
+Priority: low
+Status: formalised
+Witness: `python aggregator_profiling/profile_lens_aggregator.py --quick` run from the autolens_workspace_developer root completes without ConfigException.
+Unattended: ready
+Filed: 2026-09-08
+
+Found while smoke-testing PR autolens_workspace_developer#133 (task aggregator-search-json-sentinel, PyAutoFit#1582).
+
+`aggregator_profiling/mock_lens_results.py:85` calls
+`conf.instance.push(new_path="config", output_path=str(root))`, but the repo has no
+`config/` directory (tracked or untracked). Run as its docstring instructs, the
+script dies immediately with `autonerves.exc.ConfigException: config does not exist`.
+A bare `conf.instance.push(new_path="config", ...)` from that root fails identically,
+so the fault is the missing config tree, not the script logic. It was added in #218
+and can only ever have run from a cwd carrying a config tree. The smoke for #133 was
+run from a scratch directory symlinking `autolens_workspace/config`.
+
+Fix options: point `new_path` at the sibling `autolens_workspace/config` via a
+resolved path (as other developer scripts do), or vendor a minimal config tree.
diff --git a/tasks/quick_update_plotting_cost.md b/tasks/quick_update_plotting_cost.md
new file mode 100644
index 0000000..6018af9
--- /dev/null
+++ b/tasks/quick_update_plotting_cost.md
@@ -0,0 +1,98 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/research/autolens/quick_update_plotting_cost.md
+Migrated-at: 2026-10-03
+
+Original task retained verbatim below. Current scheduling state: `campaigns.yaml`. Historical paths: `migration.yaml`.
+
+---
+
+# Quick-update plotting cost — minutes per update, and it is not JAX compile
+
+Type: research
+Target: autolens
+Repos:
+- PyAutoLens
+- PyAutoGalaxy
+Themes:
+- visualization
+- profiling
+Difficulty: medium
+Autonomy: safe
+Priority: medium
+Status: draft
+Consequence: glance
+Witness: A breakdown document and a runnable timing harness (`perform_quick_update` cold/warm on the wst MGE-linear setup) are committed, attributing per-update cost across critical-curve extraction on the unmasked grid, `cached_property` fit evaluation, device transfers and savefig, with no numbers claimed from a loaded box; before/after per-update seconds appear only once the measurement leg has run on a quiet local box.
+Review-minutes: 3
+Unattended: ready
+Filed: 2026-07-31 (backfilled from git)
+
+## Why
+
+While removing `fit_quick.png` (PyAutoLens#680, merged 2026-07-31), a control
+probe on the wst `modeling_visualization_jit.py` MGE-linear imaging setup
+measured `perform_quick_update`:
+
+- old 6-panel `fit_quick`: cold 199 s / warm 184 s
+- new normal `fit.png` subplot: ~296 s first call
+
+Cold ≈ warm, so the cost is **recurring** per-update fit-quantity computation
+plus rendering — NOT one-off JAX compile (the `fit_for_visualization` jit
+compile is ~27 s, shared, and its warm call is ~2.6 s). Numbers are from a
+heavily loaded WSL box (load ~5), so absolute values are inflated, but a
+~3-minute recurring cost per quick update dwarfs the seconds-scale search
+iterations it interleaves with. The user expected the fit/search side to be
+seconds — it is; the plotting path is the sink.
+
+## What
+
+Profile where the time goes in `PlotterImaging.fit_imaging` /
+`VisualizerImaging.visualize` when called from the quick-update hook with a
+jit-cached fit: fit `cached_property` evaluation (inversion / per-plane model
+images / source-plane quantities), JAX→NumPy device transfers, critical-curve
+computation (`_compute_critical_curves_from_fit`), and matplotlib rendering/
+savefig. Produce a breakdown, then propose and (if warranted) implement the
+cheap wins — e.g. reusing already-computed quantities, computing critical
+curves less often, or skipping expensive panels during quick updates without
+reintroducing a second fit-plot layout (the whole point of #680 was one
+layout). Re-measure on a quiet box and record before/after per-update cost.
+
+Probe script from the #680 session (rebuildable): scratchpad
+`quick_update_timing_probe.py` — builds the wst MGE-linear analysis, times
+`perform_quick_update` cold/warm.
+
+## Triage 2026-09-09 — why this needs a local box, and what a static pass already found
+
+Considered for the `visualization` bundle and **dropped**: the deliverable is a
+*measured* breakdown plus a before/after re-measure, and the prompt itself
+discounts its original numbers because the box was loaded (load ~5). A remote
+Claude container is a worse measuring instrument than the one those numbers came
+from — shared 4-core, no lensing stack, no JAX, and no wst MGE-linear dataset —
+so any timing it produced would be exactly the untrustworthy kind this task
+exists to replace. **The measurement leg is human/local (or RAL) work.**
+
+A static read of the quick-update path on `main` (2026-09-09) did narrow the
+suspects, and these are worth checking first when the profiling is actually run:
+
+- `autolens/imaging/model/visualizer.py:108` calls
+  `_compute_critical_curves_from_fit(fit)` on **every** quick update. That
+  resolves (`autolens/imaging/plot/fit_imaging_plots.py:25`) to
+  `_compute_critical_curve_lines(tracer, fit.mask.derive_grid.all_false)` — the
+  **fully unmasked, full-resolution** image-plane grid, deliberately, so the
+  curves cover the whole image extent. Critical-curve extraction over that grid
+  is a per-update cost that scales with the *unmasked* image size rather than
+  the fitted region, and it is recurring, not compiled once. This is the single
+  strongest candidate for the "cheap win" the prompt asks for — e.g. computing
+  the curves on a coarser grid, or less often than every quick update.
+- The curves are already hoisted correctly: the visualizer computes them once
+  and passes them into `plotter.fit_imaging`, which reuses them
+  (`autolens/imaging/model/plotter.py:69-79`). So the *duplication* win is
+  already taken; what remains is the cost of the single computation.
+- `_get_source_vmax` (`fit_imaging_plots.py:38`) walks
+  `fit.model_images_of_planes_list`, a `cached_property` chain that triggers the
+  inversion — cheap if the fit is already warm, not if the quick-update hook
+  hands over a fresh fit.
+
+Suggested re-scope if this is picked up unattended: split the *analysis + an
+instrumented probe committed to the repo* (deliverable: a breakdown document and
+a runnable timing harness, no numbers claimed) from the *measurement + tuning*
+(deliverable: before/after per-update cost on a quiet box). Only the first half
+is container-safe.
diff --git a/tasks/runtime_cell_single_jit_gpu_warmup.md b/tasks/runtime_cell_single_jit_gpu_warmup.md
new file mode 100644
index 0000000..5286772
--- /dev/null
+++ b/tasks/runtime_cell_single_jit_gpu_warmup.md
@@ -0,0 +1,66 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/bug/autolens_profiling/runtime_cell_single_jit_gpu_warmup.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Runtime cells' A100 `single_jit` includes the post-compile warm-up — source-plane 0.642 ms vs a steady 0.267 ms on the same node
+
+Type: bug
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+Themes:
+- profiling
+- jax
+Difficulty: small
+Autonomy: supervised
+Priority: low
+Status: formalised
+Consequence: judge
+Witness: On `euclid-ral-gpu-2` the release-sweep runtime cells' `full_pipeline_single_jit` for the A100 row agrees with a steady-state median of individually timed calls (≥ 5 warm calls, ≥ 200 timed) to within the median's p10–p90 band — first for `point_source_source/source_plane_solved` (steady 0.267 ms, job 366914), then for each imaging cell the check covers; or, if the decision is to keep the statistic, the dashboard labels the A100 `single_jit` as "first block after compile" and headlines `vmap.per_call` instead.
+Review-minutes: 10
+Unattended: needs-decision
+Epic: point-source-cpu-speed
+Filed: 2026-09-28
+Updated: 2026-09-28
+
+## Finding
+
+autolens_profiling#349 (2026-09-28) re-ran `scripts/point_source_source/likelihood_runtime/source_plane_solved.py`
+on the release-sweep reference node at 2026.9.27.2. The A100 row (job 366912, load 1.16, cache fresh)
+recorded `full_pipeline_single_jit` = **0.642 ms**. A same-node diagnostic (job 366914) built the
+identical analysis, warmed the AOT `Compiled` executable with 5 calls, then measured:
+
+- the cell's own statistic (mean of 10 consecutive `block_until_ready` calls), 40 times: median
+  **0.272 ms** (p10 0.250, p90 0.285);
+- 400 individually timed calls: median **0.267 ms** (p10 0.260, p90 0.279);
+- scalar `jit(x + 1)` floor 0.132 ms.
+
+So the committed 0.642 ms is the cell's single 10-call block landing in the post-compile transient
+(`jit_profile`: one `first_call`, then `steady_x10`), not the likelihood. The RAL CPU row (job 366911)
+agrees with earlier medians, so this looks GPU-specific. Logs:
+`results/logs/point_source_source/point_source_source_plane_2026_09_28_ral_job_{366912,366914_timing_diag}.out`;
+ledger `results/notes/point_source_source_plane_campaign.md` "Runtime refresh on 2026.9.27.2".
+
+## Reach (noted, not measured)
+
+The four imaging release-sweep cells (`scripts/imaging/likelihood_runtime/{mge,pixelization,delaunay,delaunay_nn}.py`)
+use the same `jit_profile` (one warm call, mean of 10). Their calls are ms-scale, so a sub-ms
+transient matters less in relative terms, but nobody has measured it. The run-time dashboard plots
+`single_jit` as the per-call headline for every release-sweep series.
+
+## Decision needed
+
+Changing the statistic (more warm calls, median of N) breaks comparability with every committed row
+and the dashboard trend; keeping it means the A100 `single_jit` of light cells is not a trend
+quantity. Options: (a) add warm calls + a median field beside the existing one (new field, old one
+kept for continuity); (b) keep it and have the dashboard headline `vmap.per_call` on GPU; (c) re-base
+all release-sweep cells at the next release.
+
+## Related
+
+Source checkouts report `autolens_version` 2026.8.17.1 at the 2026.9.27.2 tag (build-time stamp),
+and 322 committed HPC JSONs carry it. The dashboard versions points by that field, so source-checkout
+rows do not separate by release. This is the same release-sweep premise, and it may deserve its own prompt.
diff --git a/tasks/search_settings_estimation_infrastructure.md b/tasks/search_settings_estimation_infrastructure.md
new file mode 100644
index 0000000..f106441
--- /dev/null
+++ b/tasks/search_settings_estimation_infrastructure.md
@@ -0,0 +1,67 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/feature/autolens_profiling/search_settings_estimation_infrastructure.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Search settings-estimation + profiling infrastructure (n_starts / batch_size / n_batch)
+
+Type: feature
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+- PyAutoFit
+Themes:
+- samplers
+- profiling
+- jax-gradient
+Difficulty: large
+Autonomy: supervised
+Priority: normal
+Status: formalised
+Consequence: glance
+Witness: An estimator, given dataset + model + device, proposes `n_starts`/`batch_size` for the MultiStart family or flags that no setting fits, populated by a profiling harness the way `vram/config.py` is for Nautilus `n_batch`; the `searches/_samplers.py` builders consult it instead of hardcoding; on the 54-parameter group cell it reproduces the measured facts (16 and 32 starts fit a 6 GB card, 64 unbatched flagged).
+Review-minutes: 3
+Unattended: ready
+Filed: 2026-07-22 (backfilled from git)
+
+Right now `n_starts`, `batch_size`, `n_batch` and vmap width are hand-set per
+cell, but the correct values depend jointly on **the data, the model and the
+device VRAM**, and vary vastly between cells. `vram/config.py` already does this
+for Nautilus `n_batch` (A100-probed per dataset/model/instrument, consumed via
+`vmap_batch_for`) — there is **no equivalent for the MultiStart gradient
+family**. Build the missing half: probe or estimate per-(cell, device) settings
+for the MultiStart searches and let the builders in `searches/_samplers.py`
+consult it instead of hardcoding.
+
+Seed evidence measured on the 4-lens + 4-source group cell (54 free params,
+8 MGE bases, autolens_profiling#82), RTX 2060 6 GB:
+
+- `n_starts` and `batch_size` are **distinct levers**: `n_starts` is a
+  *scientific* knob (multi-start coverage), `batch_size` is purely a *memory*
+  one and is numerically identical to the vmap. Both have real uses; neither
+  substitutes for the other.
+- XLA compile time scales ~**linearly with `n_starts`** unbatched:
+  16 starts = 13 min 35 s, 32 starts = 26 min 53 s (projects ~53 min at 64).
+  Linear scaling suggests XLA unrolls per-start rather than compiling one
+  kernel and mapping it — worth understanding on its own.
+- `batch_size` chunking via `jax.lax.map` bounds memory but costs a lot of
+  compile: 64 starts + `batch_size=8` exceeded 44 min still compiling, vs
+  13 min 35 s for 16 unbatched. So chunking is the wrong trade **unless no
+  workable `n_starts` fits at all**.
+- Unbatched 64 starts requests a single **4.71 GiB** allocation and OOMs a 6 GB
+  card; 32 starts peaks ~5.1 GB against a 4.9 GB budget.
+- The JAX persistent compile cache (autonerves sets `JAX_COMPILATION_CACHE_DIR`,
+  `min_compile_time_secs=1`) **does** absorb this compile — a 4.27 MB
+  `jit_call` entry is written ~2 min after compile completes — so repeat runs of
+  the same graph shape are cheap, and the shared `value_and_grad` entry should
+  be reused across optimizers. Caveat: killing a run inside that ~2 min window
+  loses the entry.
+
+Goal: an estimator that, given dataset + model + device, proposes
+`n_starts` / `batch_size` (and flags when a cell simply cannot fit), plus the
+profiling harness to populate it — the MultiStart analogue of `vram/`.
+
+Related: [[project_multi_start_gradient_v2_shipped]],
+[[project_searches_first_class_a100_findings]].
diff --git a/tasks/timing_noise_audit.md b/tasks/timing_noise_audit.md
new file mode 100644
index 0000000..126d67b
--- /dev/null
+++ b/tasks/timing_noise_audit.md
@@ -0,0 +1,52 @@
+Migrated-from: https://github.com/PyAutoLabs/PyAutoMind/blob/f21898e042be67afd86cc9c512879b3ad1c9ca9a/draft/bug/autolens_profiling/timing_noise_audit.md
+Migrated-at: 2026-10-03
+
+The original task below is retained verbatim. Current scheduling state lives in `campaigns.yaml`; historical `Status` and paths below are provenance, not a second queue. Resolve old Mind paths through `migration.yaml`.
+
+---
+
+# Audit timing tests and profiling gates for measurement noise
+
+Issue: https://github.com/PyAutoLabs/autolens_profiling/issues/362
+Issued: 2026-10-02
+Type: bug
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+Difficulty: large
+Autonomy: supervised
+Priority: high
+Status: planned
+Consequence: judge
+Review-minutes: 25
+Unattended: ready
+
+# Audit timing tests and profiling gates for measurement noise
+
+Type: bug
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+Difficulty: large
+Autonomy: supervised
+Priority: high
+Consequence: judge
+
+## Goal
+Audit all runtime-sensitive tests and the production profiling acceptance mechanisms for correct treatment of measurement noise. Inventory absolute/relative cutoffs, warmup, repeated blocks, pairing/order, sample count, clock/synchronization, host qualification, estimator uncertainty and multiple comparisons. Follow dependencies into other repos only when the inventory identifies them; split implementation into bounded phases after the audit.
+
+## Trigger and evidence
+PR autolens_profiling#361 failed test_call_accounting_covers_a_real_likelihood_call at ratio 1.031073665 against 1.031; blocks [1.0218478812434695, 1.0243047139025747, 1.0470684004725312]. Equality and coverage (99.95%) passed. The guard compares block range to the entire overhead budget, not uncertainty relative to the excess. Five local repeats were all too noisy to assert the strict cutoff. The local stack differs from CI, so they demonstrate scatter, not proof of compliance. CI log: Actions run 36985476995, job 110769406158.
+
+## Acceptance
+- Produce an inventory of timing assertions and production gates, each with a noise model, assumptions, false-positive risks and ownership.
+- Define PASS / FAIL / INCONCLUSIVE semantics: inconclusive must not silently qualify a profiling result or be reported as a measured pass.
+- Preserve correctness gates, gross-regression protection, declared practical budgets and pre-registered campaign decisions; do not merely increase tolerances or retry until green.
+- Account for small samples, non-normal/outlier timings, dependence/drift and multiple comparisons; state limits of confidence-bound methods.
+- Add deterministic synthetic-data tests for clear gains/regressions, boundaries, noisy/insufficient observations, invalid timings and severe regressions, plus controlled repeated execution where needed.
+- Reconcile test helpers and production instruments so their interpretation cannot drift. Separate cheap correctness CI from hardware-dependent performance claims where appropriate.
+
+## Original user request
+fix the noise cutoff, intake na issue to fix this long term (e.g. check all trests but also make sure the whole mechanism accounts ofr noise) and then prm and continue this task
+
+<!-- formalised by the Intake (Conception) Agent on 2026-10-02 from file:tmp/noise-audit.md -->
diff --git a/tests/test_campaigns.py b/tests/test_campaigns.py
new file mode 100644
index 0000000..ccf3380
--- /dev/null
+++ b/tests/test_campaigns.py
@@ -0,0 +1,93 @@
+"""Campaign state is distinct from timing evidence and survives board refreshes."""
+
+import copy
+import hashlib
+from pathlib import Path
+
+import pytest
+import yaml
+
+from pulse import ORGAN_ROOT, board, campaigns
+
+
+def write_ledger(root, data):
+    (root / "campaigns.yaml").write_text(yaml.safe_dump(data))
+    for t in data["tasks"]:
+        p = root / t["path"]
+        p.parent.mkdir(parents=True, exist_ok=True)
+        p.write_text("Task contract\n")
+
+
+def test_migration_keeps_original_text_and_every_task():
+    ledger = campaigns.load()
+    migration = yaml.safe_load((ORGAN_ROOT / "migration.yaml").read_text())
+    mapped = {t["destination"] for t in migration["tasks"]}
+    assert mapped <= {t["path"] for t in ledger["tasks"]}
+    assert len(mapped) == len(migration["tasks"])
+    for row in migration["tasks"]:
+        text = (ORGAN_ROOT / row["destination"]).read_text()
+        original = text.split("\n---\n\n", 1)[1]
+        assert hashlib.sha256(original.encode()).hexdigest() == row["sha256"]
+        assert migration["source_commit"] in text
+        assert row["source"] in text
+
+
+@pytest.mark.parametrize(
+    "mutation", ["duplicate", "campaign", "status", "missing", "escape", "url"]
+)
+def test_invalid_ledgers_fail_closed(tmp_path, mutation):
+    data = copy.deepcopy(campaigns.load())
+    write_ledger(tmp_path, data)
+    task = data["tasks"][0]
+    if mutation == "duplicate":
+        data["tasks"].append(copy.deepcopy(task))
+    elif mutation == "campaign":
+        task["campaign"] = "unknown"
+    elif mutation == "status":
+        task["status"] = "maybe"
+    elif mutation == "missing":
+        task["path"] = "tasks/missing.md"
+    elif mutation == "escape":
+        task["path"] = "tasks/../campaigns.yaml"
+    else:
+        data["campaigns"][0]["evidence"] = "javascript:alert(1)"
+    (tmp_path / "campaigns.yaml").write_text(yaml.safe_dump(data))
+    with pytest.raises(campaigns.CampaignError):
+        campaigns.load(tmp_path)
+
+
+def test_control_room_precedes_measurements_and_escapes():
+    data = copy.deepcopy(campaigns.load())
+    data["campaigns"][0]["title"] = "<script>alert(1)</script>"
+    for text in (
+        board.render_html([], campaign_data=data),
+        board.render_markdown([], campaign_data=data),
+    ):
+        assert (
+            text.index("Check in on all profiling work")
+            < text.index("Active campaigns")
+            < text.index("Active tasks")
+            < text.index("Profiling evidence")
+        )
+    page = board.render_html([], campaign_data=data)
+    assert "<script>alert(1)</script>" not in page
+    assert "&lt;script&gt;" in page
+    assert "field.value" in page  # copy the edited value, not a frozen default
+    assert "Select and copy the prompt above." in page
+
+
+def test_refresh_does_not_stamp_checkin_or_change_qualification(tmp_path):
+    data = campaigns.load()
+    original = (ORGAN_ROOT / "campaigns.yaml").read_bytes()
+    before = board.render_state([], updated="2026-10-03T00:00:00Z")
+    board.write([], tmp_path, updated="2026-10-03T00:00:00Z")
+    assert (ORGAN_ROOT / "campaigns.yaml").read_bytes() == original
+    assert campaigns.marker(data) in (tmp_path / "dashboard.html").read_text()
+    import json
+
+    assert json.loads((tmp_path / "state.json").read_text()) == before
+
+
+def test_missing_ledger_is_not_an_empty_success(tmp_path):
+    with pytest.raises(campaigns.CampaignError, match="missing campaign ledger"):
+        campaigns.load(tmp_path)