Affected Repositories

• autolens_profiling (primary; workspace-type repo)
• PyAutoMind (ledger only: active.md registration, prompt move)

Branch Survey

Suggested branch: claude/profiling-wiki-phase-2-b8vtjm (the session's designated branch, restarted from main after #340 merged).

Work Classification: Workspace (profiling repo; tooling + docs + file moves, no library change).
Worktree root: none — remote container clone.

Implementation Steps

1. Moves (git mv, filenames unchanged):
   • results/notes/{certified_solver_phase_c1_job350768,certified_solver_policy_phase_b_job350588,fixed_light_numba_s4b_source_job343397,fixed_light_numba_s5_source_job343398,fixed_light_numba_s5b_source_job343413,fixed_light_numba_s6_source_jobs,hst_gpu_residue_phase2_historical_job343376,hst_gpu_residue_phase2_job344635,hst_gpu_residue_phase3_job350573,hst_gpu_residue_phase4_job350651}.json → results/breakdown/imaging/.
   • results/notes/point_source_cpu_*.out (12) → results/logs/point_source_image/; results/notes/point_source_source_plane_*.out (4) → results/logs/point_source_source/. Add results/logs/README.md (what a log is, how it is cited, the RAL-only alternative).
   • Rewrite links: ledgers certified_solver_policy_phase_b_2026_09.md, certified_solver_phase_c1_lane_rate_2026_09.md, fixed_lens_light_s4_2026_09.md, fixed_lens_light_numba_memo_2026_09.md, fixed_lens_light_numba_memo_policy_2026_09.md, fixed_lens_light_numba_scaling_2026_09.md, hst_gpu_residue_phase{2_vmap,3_psf,4_logdet}_2026_09.md, point_source_cpu_campaign.md, point_source_source_plane_campaign.md; wiki pages certified_positive_solver.md, fixed_light_numba_cpu.md, hst_gpu_residue.md, point_source_image_plane_cpu.md, point_source_source_plane.md (whichever cite a moved file). clipper_campaign/ and point_source_cpu_2026_09_17_reported/ stay where they are (frozen evidence packs with their own READMEs).
2. Policy: new "Artefact policy" section in results/README.md (+ a logs/ row in the sections table): where logs, sidecars and summaries live; the summarised-JSON rule for new measurements; the provenance contract (field list, where the block sits, start/at-write loadavg semantics); the lint that enforces the notes-only rule and the provenance ratchet.
3. Provenance block: in _profile_cli.py add _LOADAVG_AT_IMPORT and provenance_dict() → {captured_at, host, slurm: {job_id, array_job_id, array_task_id}, loadavg_at_import, loadavg_at_write, profiling_revision, library_revisions (from likelihood_breakdown/provenance.source_revisions), library_versions, dependency_versions (jax, jaxlib, numpy, scipy, numba, nufftax, …), provenance_schema: 1}; device_info_dict() gains "provenance": provenance_dict(). Every failure inside it degrades to None fields, never raises.
4. Check: scripts/misc/tooling/check_results_layout.py [--check] [--write-grandfather] with the two rules above; results/provenance_grandfathered.txt generated once (pre-policy result JSONs, one path per line, header explains the ratchet: entries may only be removed). New step in lint.yml beside check_wiki.py.
5. Tests: scripts/misc/test/test_check_results_layout.py (tmp trees: clean, stray .out, stray dir, missing provenance, grandfathered) and scripts/misc/test/test_provenance_dict.py (keys, SLURM env pickup, no raise without git).
6. Validate: check_results_layout.py --check, check_wiki.py --check, build_readme.py --check, ruff check / format --check, pytest scripts/misc/test -q; open the PR; merge stays human (/prm).

Key Files

• _profile_cli.py — provenance_dict(), device_info_dict()
• scripts/misc/likelihood_breakdown/provenance.py — reused for source_revisions
• scripts/misc/tooling/check_results_layout.py — new gate; results/provenance_grandfathered.txt — manifest
• results/README.md, results/logs/README.md — policy
• .github/workflows/lint.yml — new step
• results/notes/*.md, wiki/campaigns/*.md — link rewrites only

Move committed job logs and JSON sidecars out of results/notes/

• Work type: refactor
• Target: @autolens_profiling
• Epic: profiling-research-wiki
• Autonomy: human-required
• Filed: 2026-09-27

Original prompt

From the 2026-09-27 profiling review: results/notes/ holds 74 files / 28,474 lines, of which
16 .out job logs (5,311 lines) and 10 .json sidecars (5,406 lines) are 38 % of the "notes"
tree; recent measurement PRs are 16k–26k lines each, mostly result JSON (e.g. #327 = 24,100 lines
of JSON, #335 = 15.5k). The maintainer agreed this is a problem.

Scope

1. Decide and document the artefact policy in results/README.md: .out logs go to
   results/logs/<campaign>/ (or stay on RAL and are cited by job id); JSON sidecars go beside the
   result JSONs they describe, not under notes/; committed result JSON is summarised (per-row
   medians + CI + provenance block), not per-repeat dumps.
2. git mv the existing 26 files with the ledger links rewritten (and check_wiki.py, lychee
   green). History is preserved by the move; nothing is deleted.
3. Add a lint-gate check that results/notes/ contains only .md.
4. Mandatory provenance block for every result JSON: host, loadavg at start/end, SLURM job id,
   library revisions, dependency floors (nufftax etc.) — the fields the wiki index and the future
   dashboard read.

Out of scope

Re-running anything; changing what any script measures.