Jammy2211 · Oct 13, 2026 · Oct 9, 2026 · Oct 13, 2026 · Oct 13, 2026 · Oct 13, 2026
diff --git a/campaigns.yaml b/campaigns.yaml
diff --git a/checkins/2026-10-13.md b/checkins/2026-10-13.md
@@ -0,0 +1,123 @@
+# Profiling check-in — 2026-10-13
+
+Reviewed at 2026-10-13T14:47:44Z. Previous full check-in: 2026-10-08T06:59:54Z.
+The ledger already included later 8 October events, and open Pulse PR #48 held a
+9 October task-only update. This review retains both; it does not claim those
+already recorded events as new measurements.
+
+## Verified sources and limits
+
+- [Profiling main](https://github.com/PyAutoLabs/autolens_profiling/tree/4db4d988e520a82219152743748c3188d485a4ba):
+  only results notes changed after #407; no new committed measurement JSON.
+  `gh run list --workflow profile.yml --limit 5` returned no runs. This does not
+  establish that no measurements exist outside GitHub.
+- [Current catalogue](https://github.com/PyAutoLabs/autolens_profiling/blob/4db4d988e520a82219152743748c3188d485a4ba/dashboard/catalogue.json):
+  112 indexed records, 0 qualified, 0 comparisons, generated 2026-10-09T10:50:50Z.
+  `evidence_updated_at` and `valid_until` are null: no common measurement clock
+  and no freshness policy. Render/fetch/review dates do not re-age evidence.
+- [Source routes](https://github.com/PyAutoLabs/autolens_profiling/blob/4db4d988e520a82219152743748c3188d485a4ba/catalogue/script_routes.json):
+  per-cell compile builder unavailable. Archived compile results do not make it executable.
+- Brain profiling conductor, run locally on 2026-10-13: runtime triage scanned
+  334 files, no findings; compile triage read 25 pins, no findings. These are
+  limited results over stored evidence, not proof of current-release performance.
+- [Lens release 2026.10.12.1](https://github.com/PyAutoLabs/PyAutoLens/releases/tag/2026.10.12.1),
+  [Array](https://github.com/PyAutoLabs/PyAutoArray/releases/tag/2026.10.12.1) and
+  [Galaxy](https://github.com/PyAutoLabs/PyAutoGalaxy/releases/tag/2026.10.12.1)
+  published 12 October. No matching fresh profiling sweep was verified.
+- Project driver `hpc/sync jobs` succeeded and returned an empty current queue.
+  `hpc/sync status` preview timed out at 25 seconds, after listing log changes
+  and starting the output scan. Nothing was pulled; historical jobs, unpulled
+  results and cleanup directories remain unverified. An empty queue is not completion.
+- [Mind claims](https://github.com/PyAutoLabs/PyAutoMind/blob/35281b1d/active.md)
+  reserve autolens_profiling for the three timing-audit PRs awaiting merge.
+  No Pulse claim is present. Heart helper: Release STALE; monitoring RED,
+  score 37/100, 256 unresolved. This is not a profiling regression verdict.
+
+## Campaign review
+
+| Campaign | Verified change / current state / next step |
+|---|---|
+| Setup baseline collection and scientific acceptance | Draft unchanged; no frozen revisions/environment lock, selected hardware/affinity, complete per-cell settings/input hashes/witness tolerances or collection authorization. Compile builder remains explicitly unavailable. Resolve the manifest choices, then request separate collection authorization. |
+| Point-source image plane · CPU | No new committed campaign measurements. Phases 1–4 remain shipped; epic closure is a human decision. Carried leftovers remain pending. RAL queue empty at review, but transfer preview timed out: cleanup folders and unpulled results unverified. Select one owed cleanup/measurement phase only after review. |
+| Point-source image plane · A100 | No new image-plane fit measurement found. Correctness fix #768 remains shipped; current libraries now have release 2026.10.12.1, not a measured performance baseline. Image-plane admission measurement still needs authorization. #411 proposes reinterpreting the old GPU memo MDI evidence; it remains unmerged. |
+| Point-source source plane | Core remains parked; autolens_inference#21 reorganized setup evidence but added no point-source gradient admission measurement. #408 merged median fields into 11 runtime cells; #412 completes remaining headline/repeat support but is open. No new runtime sweep; keep Blackjax parked. |
+| Cluster PointSolver robustness and performance | No campaign wiki/results change or new matching implementation claim found. Representative cluster baseline and observable-overflow policy remain unstarted. Choose one bounded phase; inherited single-source results do not qualify a representative cluster baseline. |
+| Interferometer likelihood and streaming | No new committed measurements. Decision matrix remains filled; streaming 1e8 CPU/A100 rows and the #332 memo after-measurement still need authorization. #411 proposes making the historical numba kill gate INCONCLUSIVE (four rounds); it is unmerged. Curvature preload needs the fixed-mapper science decision; FFT levers remain unstarted. |
+| Linear-solver accuracy and cost | Retain the 2026-10-08 human decision: Jacobi for Mapper, raw+polish for MGE-only, no default change. Project wiki still says decision pending (stale). No new solver measurements after the recorded phase-5 results; release 2026.10.12.1 does not supply a rerun. Wide-prior/interferometer gaps remain evidence gaps, not authorized tasks. |
+| Production imaging and likelihood breakdown | No new production-over-sampling or post-certified measurements. #408 now adds median headline fields to runtime cells; #412 extends remaining surfaces but is unmerged. Production presets, memo-blind harness, cumulative MGE steps and quick-update costs remain separate pending work. Scope production-preset phase 1 (CPU smoke only); A100 measurement separately authorized. |
+| Measurement reliability and profiling tools | #408/#409 merged 2026-10-09. #410 -> #411 -> #412 remain OPEN, lint SUCCESS at reviewed heads ffddf755/a5de4d49/9ce9a132; Mind claims all three awaiting merge. Review flagged statistical/estimator decisions, then human /prm in order. Issue #362 stays open. No new measurement JSON; v2 catalogue remains 0 qualified / 112 records, 0 comparisons. |
+| PyAutoFit profiling bootstrap | B1 registration is complete: Heart#292 and .github#34 merged 2026-10-08. autofit_profiling main remains skeleton 86c77345: no exporter, harness or measurements, no Pulse fit registration. B4a remains next profiling phase, subject to the search-extensibility epic handoff/review findings; then B4b. Do not restart the separate B3 inference pilot. |
+| Critical curves and evaluation grids | No new campaign measurement. Phase 3b cap fix remains shipped; project wiki still incorrectly says #646 awaits merge/release. Seed coverage and path completion remain unresolved; retain marching-squares decision and ask the human whether to scope the next phase. |
+| Certified positive solver | Parked unchanged. No new campaign evidence; existing C2/library/workspace drafts remain pending. Do not infer admission from solver timing, and do not change default or restart the campaign. |
+
+## Open PRs and proposed priority order
+
+1. Review the timing-audit stack, then human merge in order:
+   [#410](https://github.com/PyAutoLabs/autolens_profiling/pull/410)
+   `ffddf755`, [#411](https://github.com/PyAutoLabs/autolens_profiling/pull/411)
+   `a5de4d49`, [#412](https://github.com/PyAutoLabs/autolens_profiling/pull/412)
+   `9ce9a132`. All three open, lint SUCCESS when checked. Existing independent
+   review reports are historical evidence; this check-in did not repeat a code review.
+   Preserve the flagged choices: C1 four repeats vs five-round minimum; C5's
+   two-estimator rule; C11 four rounds vs a separately authorized rerun;
+   GPU memo readings under paired-round MDI; headline/wall/repeat choices in #412.
+   Re-analysis is not a new hardware measurement. No merge authorized here.
+2. Resolve the compile-builder gap and freeze the setup-baseline specification.
+   First choose runtime-only scope or a bounded compile-builder implementation;
+   then freeze immutable software/environment, hardware, dataset hashes, models,
+   precision, cache state, synchronization/repetitions and correctness witnesses.
+   Preparation needs engineering/review; compute is separately authorized.
+3. Plan the owed interferometer memo after-measurement on the same ALMA Delaunay
+   setup and matched hardware/settings. This would test whether the released
+   back-off fixes the original workload; the solver-only witness cannot answer it.
+   A quiet CPU allocation and explicit compute authorization are needed.
+4. Prepare production imaging presets and representative cluster baselines,
+   one bounded task at a time after the current profiling claims clear. CPU smoke
+   validates instrumentation; any CPU/A100 timing allocation is a separate phase.
+5. Coordinate autofit B4a with its existing epic; preserve parked source-plane
+   and certified-solver work. GPU admission, streaming extensions and critical-curve
+   seed/path experiments require a specific human choice before execution.
+
+## Related handoffs and stale prose
+
+- [Autofit skeleton](https://github.com/PyAutoLabs/autofit_profiling/tree/86c77345)
+  unchanged; [Heart #292](https://github.com/PyAutoLabs/PyAutoHeart/pull/292) and
+  [.github #34](https://github.com/PyAutoLabs/.github/pull/34) merged 8 October.
+  B4a remains unstarted. The [epic](https://github.com/PyAutoLabs/PyAutoMind/blob/35281b1d/draft/research/autofit/search_extensibility_epic.md)
+  records post-merge review findings and recovery history; do not restart B3's pilot.
+- [Inference #21](https://github.com/PyAutoLabs/autolens_inference/pull/21)
+  reorganized frozen setup evidence; it supplied neither a new image-plane fit
+  measurement nor a point-source gradient admission measurement.
+- Linear-solver wiki's decision-pending text is superseded by the
+  [8 October human decision](../decisions/linear-solver-mapper-default-jacobi.md).
+  Critical-curves wiki still says the cap fix is unmerged; GPU wiki still names
+  the fixed forward-gradient NaN. These project prose discrepancies are flagged,
+  not silently rewritten during a Pulse check-in.
+- All 24 open Pulse tasks reviewed; completed/superseded task states retained.
+  Pending task contracts are unchanged except for dated evidence/next-step notes.
+
+## Comparison discipline
+
+No new speedup or regression is asserted. Existing results span laptop WSL CPU,
+RAL Xeon/EPYC and A100; they cannot be pooled. Software Git revisions (including
+JAX/jaxlib), dataset/model/over-sampling/solver settings, precision, cache state,
+thread placement and measurement estimator must match before a timing claim.
+Legacy block means and new per-call medians remain distinct series. Keep cold
+compile/first-call/warm-up separate from steady runtime, solver-only timings
+separate from complete likelihood cost, and RSS separate from device memory.
+
+No jobs, releases, baseline/default changes or campaign acceptance were authorized
+or performed by this check-in. PR publication does not merge the ledger.
+
+
+## Subsequent authorized action — 2026-10-13
+
+After this check-in the human approved the priority order and explicitly said
+“Merge all four when green.” Profiling #410, #411 and #412 merged unchanged in
+order, with all exact-head Actions runs/jobs successful. Issue #362 closed;
+Mind's three records are complete and claims released. This supersedes the
+open-PR status in the review-time table above without changing that historical
+review or measurement timestamps. Pulse #48 carries this follow-up as well.
+The audit implementation is complete; the campaign remains active for its other
+tools and the deliberately unresolved domain choices. No compute or scientific
+acceptance follows from these merges.
diff --git a/dashboard.html b/dashboard.html