@@ -0,0 +1,46 @@
+# Likelihood lifecycle phase 6d: ingest the gradient witnesses, freeze day_zero_gradient, file hazards by mechanism
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+- PyAutoPulse
+Themes:
+- profiling
+- jax-gradient
+Difficulty: medium
+Autonomy: supervised
+Priority: high
+Status: formalised
+Consequence: judge
+Witness: every Optimal setup's gradient leg (74) has a current mirrored gradient.json under results/ read by build_catalogue as a measurement gradient record with an FD verdict; the 10 FD-pass records carry day_zero 2026.10.25.1; Pulse snapshots/day_zero_gradient/lens.json holds exactly those gradient records (runtime day_zero untouched); every non-pass verdict and every grad/forward > 4x is covered by a hazard prompt filed by mechanism; the epic's phase-6 witness is restated
+Review-minutes: 20
+Unattended: never
+Epic: likelihood-lifecycle
+Phase: 6
+Parent: draft/research/autolens_profiling/likelihood_lifecycle_epic.md
+Filed: 2026-11-02
+
+## Context
+
+- 6a autolens_profiling#506 (harness, gradient_witness, build_catalogue gradient records), 6b #508 (gradient_breakdown.py ×19, gradient_p6 74 legs), 6c #509 (RAL runs: v2 RAL 413791–413830, 72/74; OOM legs rerun at 512 GB as RAL 413890; witness v2 step sweep #510; walls #511 1cdf84bf).
+- v2 sweep verdicts: agree (FD pass): cluster source_plane, group mge, multi_galaxy mge, point_source_source, weak (CPU + A100 = 10 records). mismatch: imaging mge ×4, interferometer mge sma. zero: cluster image_plane, point_source_image. no_plateau: every Delaunay / delaunay_nn / rectangular cell.
+- Delaunay/Hilbert side project (RAL 413876) accepted by the human 2026-11-02 ("I accept them, no follow up"): roughness is intrinsic (triangle flips); no Hilbert setting gives a plateau.
+- RTU / CPU-sparse Bilinear fiducial wrong gradient filed separately: draft/bug/autoarray/rectangular_fiducial_gradient_wrong_rtu_and_cpu_sparse.md (human handles in another chat).
+
+## Plan (human rulings 2026-11-02)
+
+1. autolens_profiling: copy the v2 gradient.json from RAL p2dg output/p6_gradient/results into the mirrored results/ cells (73 now; imaging/delaunay/ao RAL CPU when RAL 413890_0 lands); regenerate catalogue / coverage / day-zero pages / dashboard; lint green. The 10 FD-pass records carry day_zero.
+2. PyAutoPulse: `freeze --measurement <m>` filter; `pyauto-pulse freeze --instance lens --label day_zero_gradient --measurement gradient --commit <sha>` writes snapshots/day_zero_gradient/lens.json with only gradient records; snapshots/day_zero untouched.
+3. Hazards by mechanism through /intake (human: "File by mechanism"): B wrong gradient (imaging mge ×4 instruments, interferometer mge sma); C densely non-smooth pixelizations (Delaunay, delaunay_nn, rectangular; Delaunay intrinsic per the accepted finding); D noisy likelihood (interferometer rectangular sma); E piecewise-constant image-plane point/cluster; cost hazards grad/forward > 4x (cluster source_plane CPU 11.6x, weak CPU 4.8x, multi_galaxy mge CPU 4.2x, delaunay_nn/hst CPU 4.09x). A (fixed-step false alarms) recorded, not filed.
+4. Epic: restate the phase-6 witness (human ruling 2026-11-02, "Restate witness"): phase 6 is met when every Optimal setup has a current gradient record with an FD verdict — pass, or a non-pass filed as a hazard by mechanism.
+
+## Rules
+
+No compute. Merge only by the human's /prm.
+
+## Request (human, 2026-11-02, verbatim)
+
+"prm the follow-up PR when it's green, can we keep going on the task as a whole as were currently just waiting on this one off slow run"
+
+Rulings (AskUserQuestion, 2026-11-02): phase-6 witness → "Restate witness (Recommended)"; frozen label → "day_zero_gradient (Recommended)".