Slot Review — PM

Question this member was dispatched to answer

Is the numba CPU positive-only solve the same linear algebra as the JAX sparse-operator mode — and if not, is it GPU-portable or intrinsically CPU-suited? (Also stands in for the numba-cpu-likelihood epic’s next step.)

Readout — the verdict

Same linear algebra, deliberately different machines — keep both, port nothing. Identical formulation (regularised normal equations, W = HᵀN⁻¹H); every difference is hardware-deliberate: explicit 177 MB sparse preload + branchy active-set fnnls + cross-eval warm-start memo (CPU cache/state levers) vs matrix-free FFT operator + fixed-shape PDIP with differentiable relaxed-KKT VJP (GPU/vmap/grad levers). Each transplant was already measured and failed (warm-start hurts PDIP 17→38 iters; BPP/ADMM slower/fail). Two scholar-mode intake candidates offered (a design note recording the two-modes decision; landing/retiring the ledger’s pending A100 rows).

Health evidence

Full verdict posted as a comment on PyAutoArray#513 (verified at collect); no PR by design (no library changes); Mind ledger merged. Judge tier — parks for this ruling, no decide-and-flag.

Ruling

Accept the verdict (issue closes, prompt retires) / reject with reasoning; separately: /intake either scholar-mode candidate?

Where to look

PyAutoArray#513 — the verdict comment is the deliverable.

Question this member was dispatched to answer

Plain CLI runs should not print the “You are not running in a Google Colab environment” block; notebooks keep it.

Witness

CLI run prints nothing; notebook path still surfaces it; asserted in test_setup_colab.py. Witness test present and green in CI (context-detection via live IPython shell; silent path demoted to logger.debug).

Health evidence

PyAutoNerves#157: 3/3 checks success (unittest 3.12/3.13/nojax), mergeable clean, 169/169 tests, adversary leg reported CLEAN (independent model, report in PR conversation). Verified against GitHub at collect.

Ruling

Merge (/prm 157) / tweak / reject. notify tier — witness-covered.

Where to look

PyAutoNerves#157

Question this member was dispatched to answer

queue_filing.yml red on every run since 2026-08-24; queue_actions.yml lost commits on push conflict and failed silently.

Witness

Post-merge only: re-apply queue-intake to #71 → green run + filing PR. Cannot execute pre-merge (label-triggered workflows run from main) — basis-cited instead (local red/green repro of the 30 theme-import failures; live-log confirmation of the diagnosed death step).

Health evidence

PyAutoMemory#76: validate green, mergeable clean. Three fixes (missing PyAutoBrain checkout in the gate; conflict-proof rebuild-retry replacing the discarding rebase; failure reporting). Adversary found 2 real issues, both fixed in b5bc248; finding 3 (same retry shape in two sibling workflows) recorded as follow-up. decide-and-flag: opened the PR rather than park (supervised cap) — the one flagged decision.

Ruling

Merge, then re-drive #71 and #72 (label off/on) and #69 — the witness completes post-merge (checklist in the PR). Queue the finding-3 follow-up?

Where to look

PyAutoMemory#76 (post-merge checklist at the bottom)

Question this member was dispatched to answer

Move the resampling diagnostics from mid-file to a closing “Resampling Info” section of search.summary.

Witness

File ends: blank line after Visualization time → subheader → six lines. Observed on a real smoke run and asserted by a new ordering test; searches without counters write byte-identical files (differential check vs main).

Health evidence

PyAutoFit#1554: 4/4 checks success, mergeable clean; 2328 passed/44 skipped; adversary CLEAN. decide-and-flag: opened the PR rather than park (bug-cap supervised).

Ruling

Merge first of the three PyAutoFit PRs (/prm 1554).

Where to look

PyAutoFit#1554

Question this member was dispatched to answer

MultiStartProdigy with iterations_per_quick_update=50 reportedly terminates at 49 steps.

Readout — the bug as reported does not exist

The search cannot stop at a quick-update boundary — the cadence is inert for this family (the sibling member’s EXEMPT entry). The “49” was Total Samples = 49 (1 best + 48 per-start final points, fixed at construction) misread as a step count, while the run’s real length appeared nowhere in the summary. Fix: the summary now labels the composition and reports Total Steps and Stop Reason; a new e2e regression runs a cadence mid-budget to completion and carries a quick_update_count == 0 tripwire the sibling wiring change must flip.

Health evidence

PyAutoFit#1555: 4/4 checks success, mergeable clean; witness reproduced at the reported configuration; adversary first pass found 2 issues (vacuously-green e2e; composition label false on reload paths), both fixed, re-review CLEAN.

Ruling

Accept the reframed root cause and merge second (/prm 1555).

Where to look

PyAutoFit#1555 — the investigation is on #1553.

Question this member was dispatched to answer

Add a GitHub Page link to each of the six generated organ board HTML pages (witness: link asserted in each generator’s board test).

Status

Finished 20:51Z, 5 PRs green: PyAutoBrain#326, PyAutoHeart#194, PyAutoHands#273, PyAutoMemory#77, PyAutoScientist#25. Merge/tweak/reject as one ruling — six repos, one witness.

Question this member was dispatched to answer

Certify the InferenceRefs_v1 rows: legacy-MGE reuse rulings first (surfaced as packet members), fresh RAL submissions for missing/rejected rows (incl. re-queued refs 5/6/7). Phase 2 stays queued behind these baselines.

Status — RAL job 342091, read 2026-09-01 12:55Z

Dispatched from the laptop 21:08Z: RAL job 342091 (sbatch --array=0-8,10 --requeue), config hpc_a100_fp64_ref, A100 80 GB, n_live 300 (2× fiducial), 6:00:00 wall per task. The array ran one task at a time on the single GPU and all ten finished overnight: first task started 2026-08-31 22:00:39 BST, last ended 2026-09-01 08:08:54 BST — 10 h 08 m of continuous GPU, every task COMPLETED 0:0 on euclid-ral-gpu-2, no requeue, no preemption, the queue now empty for this project. Task 9 (mge_pos) was held back under the MGE reuse rule — its ruling and the two retro-baseline rulings are the three members above. Recorded on autolens_profiling#201 with docs-only PR #202. Refreshed here by hpc/sync pull to /mnt/c/Users/Jammy/Science/inference_programme/.

Health evidence — classified from artefacts, not from sacct

Every one of the ten tasks clears all five checks, so the classification below is evidence-based rather than an ExitCode 0:0 reading:

• Wall plausible — 00:36:21 to 01:24:16, i.e. 10–23% of the 6 h budget. Nothing near the wall, nothing suspiciously instant.
• Result row written — ten result JSONs under results/searches/nautilus/imaging/…/hst/, each stamping version 2026.8.17.1, each with a target_id matching its intended cell.
• .err clean — all ten error files are 222 bytes and contain only the mask_2d_util.py:564 UserWarning: Mask padded from (141, 141) to (159, 159). No Traceback in any error or output log.
• Ended properly — every .out closes with the timing block, the results-JSON path, Finished. and a timestamp; ten .completed markers, one per run directory.
• No resume, no overflow — Fit Already Completed appears zero times across all ten logs (every search started cold), and no log_l anywhere resembles the pre-fix 3e+303 signature. The only hit from an inf|nan|e+30 sweep is the literal string model.info.

Witness — per-task readout, all ten rows final

The knn row is now a real run. At the 22:53Z refresh the mirror’s knn/hst/hpc_hpc_a100_fp64_ref.json still read the pre-fix 30,077.028 — a git stash artefact, not a result. Task 7 has now replaced it with a genuine 342091 row at 30,062.612 (logZ 29,995.668), which is 14.4 nats lower still and sits ~494 nats below the Prodigy arm bar 30,557.03. The stale-row explanation is therefore closed, and the knn cell’s underperformance is real rather than an artefact — a result for the programme to absorb, not a defect in this wave.

Shape of the wave, offered as observation only. The four positions-on arms (tasks 2, 4, 6, 10) all land within 25–183 nats of truth, with task 4 (slam_source_pix, auto threshold) the only row above it at +25.50. The positions-off pixelized rows sit 495–1,686 nats below truth, pixelization lowest. Each pos-on log carries the arm banner — POSITIONS ARM (target_class 3), Idealised: truth-derived positions, not re-solved from a completed search, PositionsLH ACTIVE — so the arms are idealised by construction and the gap to the pos-off rows should be read with that in mind.

Relevant to a ruling already on this page, stated as fact and nothing more: the delaunay-fp64 retro-baseline question was framed as “add array task 11 now, or defer until the 342091 wave lands?”. The wave has landed and the GPU queue for this project is empty.

Where to look yourself

Local checkout

/home/jammy/Code/PyAutoLabs/autolens_profiling — hpc/sync jobs / hpc/sync pull drive RAL; hpc/sync.conf points pulls at the mirror below rather than at the checkout.

Laptop mirror (everything pulled by this refresh)

/mnt/c/Users/Jammy/Science/inference_programme/ — results/searches/nautilus/imaging/<cell>/hst/hpc_hpc_a100_fp64_ref*.json for the ten rows, logs/output/output.342091_<task>.out and logs/error/error.342091_<task>.err for the logs, output/searches/… for the run directories and their .completed markers.

The two rows worth opening first

knn/hst/hpc_hpc_a100_fp64_ref.json (the replaced row) and slam_source_pix/hst/hpc_hpc_a100_fp64_ref_pos_tauto0.2_f1e8.json (the one above truth).

Live job state

ssh euclid_jump then sacct -j 342091 --format=JobID,State,ExitCode,Elapsed,Start,End; squeue -u jnightin now shows nothing for this project.

Programme record

autolens_profiling#201 (queue anchor #200) and docs-only PR #202.

No ruling chip on this page

This member carried into the slot without a review control and keeps it that way — under carry-forward it is formally ruled in the next batch. What has changed is that there is nothing left to wait for: the wave is complete, all ten rows are certified-clean by the checks above, and the readout can be judged whenever you want to. The three legacy-reuse rulings it routed into this slot (mge-pos-ref-reuse, mge-fp64-retro-baseline, delaunay-fp64-retro-baseline) do carry their own chips and are yours to rule now.

Follow-ups this refresh proposes

• [autolens_profiling] Fold the ten certified rows into the InferenceRefs_v1 table and close out #201; the knn row now needs a note that its low value is a measured result, not the old stash artefact.
• [autolens_profiling] Phase 2 was queued behind these baselines (ordered phases). With Phase 1 complete, it is dispatchable at the next slot.
• [autolens_profiling] The wave used 10 h of a single A100 for 10 tasks at 10–23% of a 6 h wall each — the per-task wall can be cut substantially, which matters if Phase 2 queues a wider array.