Decisions (recorded as docs/schema_decisions.md 53–56)

• 53 — provenance header. Migrated-from: joins PHASE_KEYS and RULING_KEYS (one
  tuple edit each + REFERENCE + tests); value = the Mind path or ledger anchor the file was
  transcribed from. The prompt's requirement stands; the current schema would have flagged
  it as drift. Batch records may carry - migrated-from: (no key check there).
• 54 — intra-Cortex sequencing. Gates: stays GitHub-refs-only. A phase that waits on
  another Cortex phase's ruling is State: planned with a body line Ready when: <phase> accepted (ruling R-…) under ## Question; the human moves it to ready after the
  ruling. A phase whose dev gate has no issue yet is planned with Ready when: <Mind prompt> is issued — add its ref to Gates: and move to gated.
• 55 — Cortex-spawned dev follow-ups get an issue at filing without leaving draft/.
  The Mind prompt carries a body line Issue: <url> (opened <date> as a Cortex gate ref; reuse in start_dev — never open a second); create_issue reuses it. Mind
  REFERENCE.md gains the rule. Same mechanism gives the 3b gate its ref.
• 56 — legacy-born encodings. Pre-Cortex runs are legacy (reusable) or legacy_wrong
  (quarantined) run lines with where:; a phase whose runs are all legacy_wrong is ruled
  drop from ready (it cannot be pulled). A programme-wide directive (the REWIND) is
  one ruling on a synthetic phase inference_programme/rewind_2026_08_31 whose runs are
  the two ids the directive names, with the directive verbatim as the ruling body. One
  phase per lens for the subhalo project. Transcribed batch records are rewritten in the
  Cortex grammar (member slug = phase slug), the Mind originals stay intact with a
  - moved-to: line — a verbatim history must live somewhere live (deviation from the
  prompt's "stub"); dev-only members of the am slot are noted, not transcribed.

The migration map (what lands in the Cortex)

phases/euclid/ (project euclid) — from draft/research/euclid/…, draft/feature/euclid/resimulate…

phases/inference_programme/ (project inference_programme)

phases/subhalo_validation/ (project subhalo_validation) — one phase per lens

phases/slope_hierarchy/ and phases/ic50_workspace/ (graphical-ep)

• slope_hierarchy/1 n25_scale_up from draft/research/graphical_ep/slope_hierarchy_n25_scale_up.md (all science) — planned, Epic: graphical-ep, Ready when: graphical-ep phase 2 (Mind) is issued — add its ref.
• ic50_workspace/1 ep_scale_up from draft/research/graphical_ep/ic50_ep_scale_up.md — planned, same; the dev half is the existing Mind prompt feature/autofit/ep_lbfgs_jax.md (named in Where to look). ep_campaign.md phase map rows 3/4 point at the Cortex; its malformed header (Status: wrap) fixed.
• Cortex epics.md ## graphical-ep + Mind cortex-half:. Project rows stay dormant (a planned phase on a dormant project is legal).

phases/inference_programme/11 cluster_extended_source_inference (cluster-strong-lensing phase 11)

From draft/research/autolens/cluster_extended_source_inference.md — planned, Epic: cluster-strong-lensing, Ready when: cluster phase 10 (Mind draft) is issued — add its ref. Cortex epics.md ## cluster-strong-lensing + Mind cortex-half:; the arc ledger's
line 11 points at the Cortex. (Its "Phase: 10" collision with cluster_regime_narrative.md
is noted in the ledger, not fixed here.)

Batch records and reviews (Cortex)

• batches/2026-08-31-am.md — transcribed: dispatched: (board opened) 2026-08-29T22:00Z,
  review-at: 2026-08-31T13:00Z, members = the 7 science members as <phase slug>: <phase path> — <runs> — <minutes> — <state>, - migrated-from:, notes: names the dev
  retrospective euclid-phase2-pr46 as Mind-only. reviews/2026-08-31-am.md in the Cortex
  grammar (HEALTH per the original; decision: drop for the four rejected, accept,
  leave-to-finish, (none) + ruled: no for the carried one; notes verbatim).
  packets/2026-08-31-am.html copied (archive of record from now on; the Mind's stays at
  its Pages URL for localStorage — both noted).
• batches/2026-08-31-pm.md — new: members inference_refs_v1_redo + the four subhalo
  phases + the three ruling phases; refreshed: lines from the Mind record's three
  refreshes; reviewed-at: 2026-09-01T15:13 (transcribed); review: batches/reviews/2026-08-31-pm.md
  (Cortex grammar; the three rulings ruled: yes, the carried members (none)).

Mind-side fixes

• queue.md feat: pyauto-status shell function for cross-repo drift dashboard #3 → retired with - moved-to: PyAutoCortex phases/subhalo_validation/…; Dashboard: flag non-main branches and print follow-up commands #9
  gains note: superseded by cortex-birth (phase 5 retires it).
• epics.md jax-compile-stall: strike the "follow-up filed under draft/research/ci/"
  sentence → "follow-up never filed (the resume door was never written)".
• Two new drafts with issues at filing (decision 55): draft/maintenance/autolens_profiling/legacy_point_output_sweep.md
  (laptop action; body from DECISIONS.md:1553-1559) and
  draft/feature/autogalaxy/ell_comps_joint_disk_constraint.md (from PROGRAMME W10 + the
  8B crash channel PyAutoFit#1535); issues on autolens_profiling and PyAutoGalaxy.
  R-20260901-03's Follow-ups: cites prompt: shelve euclid-eceb-editorial-revision — untrackable, retired from the board #203 and the sweep issue.
• batches/2026-08-31-pm.md gains reviewed-at: 2026-09-01T15:13 (transcribed),
  review-minutes-actual: (not given), review: batches/reviews/2026-08-31-pm.md; that
  review file written: science members transcribed from DECISIONS.md + the record; dev
  members UNREVIEWED with note "merged via /prm 2026-09-01; no packet ruling recorded".
• batches/reviews/AGENTS.md:24 → merge | tweak | reject | defer # science members (Cortex vocabulary): accept | rerun | drop | leave-to-finish; packets/TEMPLATE.md:64 →
  dev: merge|tweak|reject|defer|structure-ok|UNREVIEWED · science: accept|rerun|drop|leave-to-finish.
• REFERENCE.md: the decision-55 rule (Issue: line on Cortex-spawned drafts; create_issue
  reuses it) — and PyAutoMind/skills/create_issue/SKILL.md one sentence.
• Moved prompts: git rm in Mind (history stays), the Cortex files carry Migrated-from:.

autolens_profiling (PR) and the science ledgers (main commits)

• DECISIONS.md: one appended entry 2026-09-01 — Rulings of record (Cortex backfill)
  mapping each 2026-08-31+ decision to its R- id (never editing the originals); CORTEX.md
  gains the id list.
• subhalo wiki/project/results_summary.md + state.md + journal: fill the landed witness
  rows (facts), cite R-20260831-06/07, note 342093/342095 await the human. Also fix the
  broken inline comment on the .gitignore sources/ line (the user's "next sweep" —
  cheap while here). Commit on main, push.

Execution (Opus, sequential where refs are needed)

• Issue via /create_issue; prompt → active/; worktree cortex-migration for
  PyAutoCortex + PyAutoMind + autolens_profiling; subhalo edited in place.
• Slice B first (Mind): the three issues (3b gate, sweep, ell_comps) + all Mind-side
  edits + the Mind pm review + git rm of the moved prompts + stub record + epics/queue/
  vocab/REFERENCE; lifecycle.py check OK; dashboard regen; local commit.
• Slice A (Cortex): schema keys + decisions 53–56, all phase files (via cortex.py new
  • move + rule --today with rule_phase(now=) from Python for the 15:13 stamps),
    both batch records + reviews + packet copy, epics.md entries; cortex.py check OK;
    pytest green; dashboard re-render (Brain main); local commit.
• Slice C (laptop): DECISIONS correction entry + CORTEX.md (worktree autolens_profiling,
  PR) and the subhalo wiki citations + gitignore fix (main, push).
• Ship: Cortex PR → Mind PR (ledger + docs; feature-branch PR since REFERENCE/skills are
  code) → autolens_profiling PR; /prm under the standing authorisation.

Verification

• Prompt witness: lifecycle.py check OK; cortex.py check OK; Mind dashboard Epics shows
  euclid-dr1-prep with phases 0–4, 8, 9 only (3a/3b renumbered) and the entry's
  cortex-half: (rendered? no — the Mind renderer ignores it; the epic ledger's link is the
  visible half — say so); Cortex dashboard shows euclid phases 4 gated on euclid#48 +
  euclid#<3b> and 5–7 planned; every DECISIONS.md entry from 2026-08-31 cites an R- id
  that resolves (the correction entry).
• No science prompt remains under PyAutoMind/draft/research/ or active/ for the moved
  set; git log --follow on each Cortex phase names its Migrated-from:.
• Cortex: pytest green; the pm/am reviews parse; gates --grade (dry) on euclid 4 reports
  PyAutoScientist 3b-3: firewall allowlist rider (clone conductor) #48 cleared and the 3b issue open → stays gated.

Branch Survey

Suggested branch: feature/cortex-migration (worktree root ~/Code/PyAutoLabs-wt/cortex-migration/)

Cortex phase 4 — migration: split the epics, backfill the rulings, move the science batch records, fix the stale pointers

Type: feature
Target: pyautocortex
Repos:

• PyAutoCortex
• PyAutoMind
• autolens_profiling
  Themes:
• mind-workflow
• dashboard
  Difficulty: large
  Autonomy: supervised
  Priority: high
  Status: draft
  Consequence: judge
  Witness: lifecycle.py check and cortex.py check both exit 0; the Mind dashboard's Epics section shows euclid-dr1-prep with phases 0–3b, 6c, 7 only and a "cortex half" link; the Cortex dashboard shows euclid-dr1-prep phases 4, 5, 6a, 6b gated on euclid#48 + the 3b issue; every ruling in DECISIONS.md since 2026-08-31 cites an R- id that resolves
  Review-minutes: 30
  Unattended: needs-slicing
  Epic: cortex-birth
  Phase: 4
  Parent: draft/feature/pyautocortex/cortex_birth_epic.md
  Filed: 2026-09-01

Phase 4 of 7 in the PyAutoCortex birth epic. Gates: phases 2 and 3. Gates
phase 6. Slices if needed: (a) epics + prompts, (b) rulings backfill, (c) batch
records + stale pointers.

Nothing here changes science state — it relocates the record of it. Every move
is a git mv-shaped edit with the source path recorded in the destination's
header (Migrated-from:), so the Mind's history and the Cortex's join up.

Task

(a) Split the epics — the migration map from the epic ledger

Rules: Cortex phase numbers are distinct integers within a project (euclid 5
stays 5; 6a/6b become 6 and 7 — record the renumbering in both ledgers). The
Mind keeps its Phase: integers but fixes its own collisions (3a/3b → 3/4,
6c/7 → 8/9 — or whatever preserves order; document it). The euclid science
prompts' Autonomy: safe / Unattended: ready headers are wrong (ledger:
"supervised, never an autonomous ship gate") — they die with the move; the
Cortex phase files have no autonomy header at all. Mind epics.md entries gain
- cortex-half: PyAutoCortex/epics.md#<slug>; Cortex epics.md entries carry
- mind-half:. The Mind's active.md subhalo entry closes with a pointer to
the Cortex phase (no complete/ record — it never had an issue; write a
one-line complete/2026/09/subhalo-followup-moved-to-cortex.md so
lifecycle.py check stays clean).

Stays in Mind (ruled): euclid 0–3b, 6c, 7; every dev phase of every epic; the
numba-vs-jax-sparse verdict; the pyautoreduce local-data research prompts
(default: stay — decide and record).

(b) Backfill the rulings

• Every ruling in autolens_profiling/results/notes/inference/DECISIONS.md
  from 2026-08-31 onward (the REWIND entry, the batch-am rulings, the
  2026-09-01 pm rulings) becomes a Cortex ruling R-… with the human's words
  verbatim; the DECISIONS entry gains a trailing Ruling of record: R-… line
  (append a correction entry — never edit the original text, per its own rule).
  Earlier entries are commentary and stay untouched.
• The subhalo rulings (pl_sersic_0 ACCEPTED 2026-08-31, pl_eff
  leave-to-finish, the 2026-08-31 tweak rulings) become R-… entries; the
  journal and results_summary.md cite them.
• The three pm legacy-reuse rulings (mge-pos-ref-reuse,
  mge-fp64-retro-baseline, delaunay-fp64-retro-baseline REJECTED) become
  rulings with the legacy/legacy_wrong run states recorded on the phase.

(c) Move the batch records and fix the stale pointers

• PyAutoMind/batches/2026-08-31-am.md + packets/2026-08-31-am.html +
  reviews/2026-08-31-am.md move to the Cortex (an all-science slot; the one
  dev retrospective member euclid-phase2-pr46 stays listed with a note). The
  Mind keeps a one-line stub record pointing across so ledger_merge and the
  packet archive's stable-path rule are not broken: the HTML stays at the Mind
  Pages URL (localStorage notes are keyed by path) and the Cortex copy is the
  archive of record from now on.
• batches/2026-08-31-pm.md: the two carried science members and three ruling
  members are recorded in a new Cortex batch record 2026-08-31-pm.md; the
  Mind record keeps its 9 dev members and gains the missing reviewed-at:,
  review-minutes-actual: and review: fields; write
  PyAutoMind/batches/reviews/2026-08-31-pm.md from the DECISIONS.md
  rulings (dated as the human's review, 2026-09-01 15:13, marked
  "transcribed") so the citation resolves.
• Fix: queue.md feat: pyauto-status shell function for cross-repo drift dashboard #3 path (draft/research/subhalo_validation/… → the Cortex
  phase; then retire the entry — science leaves the Mind queue); queue.md Dashboard: flag non-main branches and print follow-up commands #9
  (superseded by this epic — phase 5 retires it; leave a note now);
  epics.md jax-compile-stall's nonexistent draft/research/ci/ follow-up
  (file it or strike the sentence); DECISIONS.md's nonexistent
  draft/maintenance/autolens_profiling/legacy_point_output_sweep.md and
  PROGRAMME W10's nonexistent
  draft/feature/autogalaxy/ell_comps_joint_disk_constraint.md (file both as
  Mind drafts with their issues created — they are Cortex-spawned dev
  follow-ups, and the gate rule says the ref exists at filing);
  batches/reviews/AGENTS.md:24 and packets/TEMPLATE.md:64 collapse to the
  one Cortex vocabulary for science members (dev vocabulary untouched).
• Regenerate both dashboards (pyauto-brain intake dashboard --apply,
  pyauto-brain cortex dashboard --apply); Brain is already merged (phase 2).

Acceptance

• The witness above; git log --follow on every moved file reaches its Mind
  history via the Migrated-from: header; no science prompt remains under
  PyAutoMind/draft/research/ or active/.
• PRs: PyAutoCortex, PyAutoMind, autolens_profiling (DECISIONS correction
  entries + CORTEX.md), subhalo_validation (journal citations; commits on main).

Out of scope

• Rulings before 2026-08-31 (commentary; not backfilled).
• The batch conductor (phase 5). This phase moves records; it does not change
  how the next batch is planned.
• Any RAL action. legacy_point/ relocation directed 2026-09-01 is a laptop
  action already queued in the science ledger, not part of this phase.