Jammy2211 · Oct 8, 2026 · Oct 8, 2026
diff --git a/campaigns.yaml b/campaigns.yaml
@@ -61,9 +61,12 @@ campaigns:
   next: 'Phase 3a SHIPPED 2026-10-07: autolens_profiling#394 merged 21:18Z (a55dcacb), issue #393 closed. A100 job 397475
     from a PRIVATE 2026.10.7.1 checkout (shared mirror untouched, human rule): released pdip_raw CPU vs A100 parity
     max |d flux_inactive_rel| 5.4e-14, iterations identical 81/81; still inadmissible under the pre-registered rule on
-    both devices. Next: phase 3b GPU/vmap timing cell (Mind draft linear_solver_phase3b_gpu_timing_cell.md, committed
-    208bac7b) via start_dev; awaits plan approval and A100 compute authorization. 2026-10-08: RAL euclid_dr1 arrays
-    running again on the shared base (397467/397469-74), so phase 3b stays on the private checkout.'
+    both devices. Phase 3b PR OPEN 2026-10-08: autolens_profiling#396 (issue #395), A100 job 398249 (euclid-ral-gpu-2,
+    0:38) from the private base; pdip_raw per-eval min ms CPU 1.35/0.88/0.68 vs A100 3.95/0.58/0.19 at B=1/16/50
+    (SLaM distinct lanes); jacobi 1.94x pdip_raw at A100 B=50 because one diverging lane holds the batch at the 50 cap;
+    compile 0.3-1.2 s separate. New: jacobi batched vs unbatched trajectories differ on the A100 (19/50 lanes), not
+    on CPU. Timing is not admissibility; no pin moves. Next: human /prm on #396, then decide whether phase 3 closes
+    the programme.'
   evidence: https://github.com/PyAutoLabs/autolens_profiling/blob/719459fefb40c7e5be157f652818dbf923830557/wiki/campaigns/linear_solver_accuracy.md
 - id: imaging
   title: Production imaging and likelihood breakdown
@@ -206,11 +209,11 @@ tasks:
   priority: medium
   path: tasks/mge_nnls_fix_pyautoarray_571_slam_60.md
   updated: '2026-10-08'
-  issue: https://github.com/PyAutoLabs/autolens_profiling/issues/393
-  next: 'Phase 3a (parity) shipped 2026-10-07 in autolens_profiling#394; Mind record complete/2026/10/linear-solver-p3a-a100-parity.md.
-    Remaining: phase 3b timing cell (single/vmap16/vmap50, jacobi vs raw, compile separate) from the private RAL base
-    /mnt/ral/jnightin/PyAuto_wt/linear-solver-p3 (five library clones present 2026-10-08) — Mind draft committed 208bac7b;
-    start_dev awaits plan approval; A100 rows need compute authorization. Pulse ledger PR #34 open, green, unmerged.'
+  issue: https://github.com/PyAutoLabs/autolens_profiling/issues/395
+  next: 'Phase 3a shipped (autolens_profiling#394). Phase 3b issued 2026-10-08 as #395 and delivered as PR #396 (CPU rows
+    from tag worktrees + A100 job 398249 from the private base, RAL worktree autolens_profiling_wt/linear-solver-p3b);
+    lint pending at review time. Remaining: human /prm on #396; Mind record; then a programme-level decision (phase 3
+    was the last filed phase).'
   evidence: https://github.com/PyAutoLabs/PyAutoArray/pull/595
 - id: autofit_profiling_bootstrap
   campaign: autofit

diff --git a/dashboard.html b/dashboard.html
diff --git a/dashboard.md b/dashboard.md
@@ -1,6 +1,6 @@
 # PyAutoPulse — profiling dashboard
 
-<!-- pulse-campaigns:ff3d243f508375d08397bf6b7392f3d4541d3d6aac5945ae7d005910b852d09f -->
+<!-- pulse-campaigns:ef8f9f7b3e8c8d71d8d9b57e5afc8f9293665f9c7c18c1e471989fcbc60b3215 -->
 ## Profiling Check In
 
 <details><summary>Full check-in prompt</summary>
@@ -108,13 +108,13 @@ Reviewed 2026-10-07. Decision matrix complete: last cell merged (autolens_profil
 
 <details><summary>Linear-solver accuracy and cost — active</summary>
 
-Reviewed 2026-10-08. Phase 3a SHIPPED 2026-10-07: autolens_profiling#394 merged 21:18Z (a55dcacb), issue #393 closed. A100 job 397475 from a PRIVATE 2026.10.7.1 checkout (shared mirror untouched, human rule): released pdip_raw CPU vs A100 parity max \|d flux_inactive_rel\| 5.4e-14, iterations identical 81/81; still inadmissible under the pre-registered rule on both devices. Next: phase 3b GPU/vmap timing cell (Mind draft linear_solver_phase3b_gpu_timing_cell.md, committed 208bac7b) via start_dev; awaits plan approval and A100 compute authorization. 2026-10-08: RAL euclid_dr1 arrays running again on the shared base (397467/397469-74), so phase 3b stays on the private checkout.
+Reviewed 2026-10-08. Phase 3a SHIPPED 2026-10-07: autolens_profiling#394 merged 21:18Z (a55dcacb), issue #393 closed. A100 job 397475 from a PRIVATE 2026.10.7.1 checkout (shared mirror untouched, human rule): released pdip_raw CPU vs A100 parity max \|d flux_inactive_rel\| 5.4e-14, iterations identical 81/81; still inadmissible under the pre-registered rule on both devices. Phase 3b PR OPEN 2026-10-08: autolens_profiling#396 (issue #395), A100 job 398249 (euclid-ral-gpu-2, 0:38) from the private base; pdip_raw per-eval min ms CPU 1.35/0.88/0.68 vs A100 3.95/0.58/0.19 at B=1/16/50 (SLaM distinct lanes); jacobi 1.94x pdip_raw at A100 B=50 because one diverging lane holds the batch at the 50 cap; compile 0.3-1.2 s separate. New: jacobi batched vs unbatched trajectories differ on the A100 (19/50 lanes), not on CPU. Timing is not admissibility; no pin moves. Next: human /prm on #396, then decide whether phase 3 closes the programme.
 
 [Campaign evidence](https://github.com/PyAutoLabs/autolens_profiling/blob/719459fefb40c7e5be157f652818dbf923830557/wiki/campaigns/linear_solver_accuracy.md)
 
 ### Active tasks
 
-- [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) — active: Phase 3a (parity) shipped 2026-10-07 in autolens_profiling#394; Mind record complete/2026/10/linear-solver-p3a-a100-parity.md. Remaining: phase 3b timing cell (single/vmap16/vmap50, jacobi vs raw, compile separate) from the private RAL base /mnt/ral/jnightin/PyAuto_wt/linear-solver-p3 (five library clones present 2026-10-08) — Mind draft committed 208bac7b; start_dev awaits plan approval; A100 rows need compute authorization. Pulse ledger PR #34 open, green, unmerged.
+- [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) — active: Phase 3a shipped (autolens_profiling#394). Phase 3b issued 2026-10-08 as #395 and delivered as PR #396 (CPU rows from tag worktrees + A100 job 398249 from the private base, RAL worktree autolens_profiling_wt/linear-solver-p3b); lint pending at review time. Remaining: human /prm on #396; Mind record; then a programme-level decision (phase 3 was the last filed phase).
 
 </details>
 

diff --git a/tasks/mge_nnls_fix_pyautoarray_571_slam_60.md b/tasks/mge_nnls_fix_pyautoarray_571_slam_60.md
@@ -37,6 +37,15 @@ The original task below is retained verbatim. Current scheduling state lives in
 - RAL (one probe, 2026-10-08): euclid_dr1 arrays 397467 and 397469–397474 RUNNING/PENDING on `ral` against the shared base, so the 2026-10-07 "0 jobs" window has closed; the private base `/mnt/ral/jnightin/PyAuto_wt/linear-solver-p3/` still holds the five library clones. `/mnt/ral/jnightin/autolens_profiling/results/lens/solver/` does not exist — where job 397475 wrote its JSON on RAL is unverified (the rows are committed in #394).
 - Next (unchanged): start_dev for phase 3b; A100 rows need separate compute authorization. Nothing here authorizes a submit.
 
+### Execution 2026-10-08 — phase 3b delivered (PR open)
+
+- Human approved the plan and one A100 submit in chat. Issue [autolens_profiling#395](https://github.com/PyAutoLabs/autolens_profiling/issues/395); PR [#396](https://github.com/PyAutoLabs/autolens_profiling/pull/396) open (commits b8911d8, e68448c, d7b6e1e), lint pending at review time. Mind `active/linear_solver_phase3b_gpu_timing_cell.md`, worktree `linear-solver-p3b-gpu-timing`.
+- New cell `scripts/lens/solver/timing.py` (jit(vmap) over `_solvers.batched_kernel`; SLaM batches = distinct fixture+slam48 systems, euclid lanes tiled). A100 job 398249 on euclid-ral-gpu-2 (0:38) from the private base; RAL sibling worktree `/mnt/ral/jnightin/autolens_profiling_wt/linear-solver-p3b` created because the p3 one held untracked (byte-identical) 3a artefacts. Shared mirror untouched.
+- Per-evaluation min ms (7 interleaved rounds, fp64, tag 2026.10.7.1 SHAs verified in both JSONs): pdip_raw SLaM CPU 1.353/0.881/0.682, A100 3.948/0.581/0.190 at B=1/16/50; pdip_jacobi SLaM CPU 2.488/1.630/1.396, A100 6.715/1.149/0.369. Compile walls 0.3–1.2 s recorded separately. Batched pdip_raw matches unbatched on every lane (|Δ flux_inactive_rel| ≤ 1.4e-14).
+- Slowest-lane mechanism confirmed: jacobi batch max iterations sit at the 50 cap when any lane diverges (A100 B=16/50), costing 1.94x pdip_raw at A100 B=50. New, unexplained: jacobi batched vs unbatched trajectories differ on the A100 (19/50 lanes iterations, 13/50 flags), identical on CPU.
+- Caveats: euclid lanes tiled (throughput only); A100 used the warm shared JAX compile cache; laptop under background load (loadavg ~1.9); CPU batched run on jax 0.10.2 (Nerves-excluded, no deadlock seen). Verdict in the ledger: a timing is not admissibility; no pin moves.
+- Next: human /prm on #396 → Mind completion record → Pulse task status; then the programme-level decision.
+
 ---
 
 # Linear-solver programme phase 3: GPU/vmap/A100 timing and parity rows for the solver corpus