Affected Repositories

• PyAutoBrain (primary — the cortex conductor)
• PyAutoCortex (docs only: REFERENCE.md, docs/schema_decisions.md)

Branch Survey

Suggested branch: feature/cortex-science-drift

Implementation Steps

1. agents/conductors/cortex/_cortex.py: drift_findings(projects, science_root) -> list[Finding] (kind, key, path, message) and science_root_from(projects) deriving the root from <root>/<category>/<project> local_paths. Kinds: unregistered, missing, category_status, ral_category (warning). Status rules: active/development ↔ active|dormant|planned; complete/archive ↔ retired|dormant.
2. drift verb: --json, --science-root; exit 0 clean / 1 drift / 2 bad args; root absent → skip notice, exit 0.
3. One-line summary in census and the checkin read-out (laptop only).
4. Document in the conductor AGENTS.md, PyAutoCortex REFERENCE.md (Check-in section) and a dated docs/schema_decisions.md entry.
5. Tests in tests/test_cortex_conductor.py with a tmp Science tree: each finding kind, exempt Code rows, skip notice, --json.
6. Validate on the real tree: expect only the four development projects' ral_category warnings (RAL trees still under active/).

Key Files

• PyAutoBrain/agents/conductors/cortex/_cortex.py — conductor (census, checkin, new verb)
• PyAutoBrain/tests/test_cortex_conductor.py — tests
• PyAutoCortex/REFERENCE.md, PyAutoCortex/docs/schema_decisions.md — docs

Cortex as the Science project manager: drift check and project birth

Type: feature
Target: cortex
Repos:

• PyAutoBrain
• PyAutoCortex
  Themes:
• science
• cortex
  Difficulty: medium
  Autonomy: human-required
  Priority: high
  Status: formalised
  Consequence: judge
  Filed: 2026-10-29

Original request (verbatim)

"5) Bring this all together with PyAutoCortex which needs to become the main tool to manage Science
projects, its meant to be but hasnt been used as such for a while."

Follow-up (2026-10-29), after the Science clean-up session proposed next steps
(1 push the path-fix PR, 2 a drift check comparing the Science folders to the register,
3 a new-project command that scaffolds a project in active/ or development/, seeds it from
dataset/ or simulators/ and registers it in one step): "yeah do it".

Context

The Science folder (/mnt/c/Users/Jammy/Workspace/Science) was reorganised on 2026-10-29:
development/ holds software-development projects, simulators/ holds the standard simulated
lenses (shared --dataset-path / --dataset-types CLI, default output dataset/simulated/),
profiling/ holds RAL pull mirrors with a generated INDEX.md, and Science/AGENTS.md names
the Cortex as the project register. Before that, the move from C:\Users\Jammy\Science to
Workspace\Science had left most projects.yaml local_paths pointing at nothing, unnoticed for
weeks: pull --dry-run reached 3 of 12 active projects. PyAutoCortex#77 fixes those paths and
retires subhalo, cowls, cowls_ansofi; it lands before this work.

The failure mode to prevent: the register and the Science tree silently diverge. And starting a
project today is manual copying from an older project, so new projects never get a row.

Scope

Phase 1 — drift check (conductor verb, report-only)

pyauto-brain cortex drift [--science-root PATH] [--json], also summarised in census and
checkin when the Science root is on this machine (laptop-only, like pull; elsewhere a one-line
skip notice, never a failure). It writes nothing. Findings:

1. a project folder under active/, development/, complete/, archive/ with no row
   (matched by local_path);
2. a non-retired row whose local_path does not exist;
3. category vs status: rows under active/ or development/ should be active|dormant|planned;
   under complete/ or archive/ should be retired|dormant;
4. (warning) a non-retired row whose ral_root category differs from its local_path category
   (decision 61: one layout on both machines).

The Science root is derived from the local_paths (<root>/<category>/<project>), overridable
with --science-root; no new absolute path enters the organism outside projects.yaml.
Rows under ~/Code/PyAutoLabs (repo-backed campaign roots) are exempt from 1/3/4.

Phase 2 — project birth (conductor verb)

pyauto-brain cortex birth <key> --category active|development --summary "<one line>" [--dataset <sample>[/<lens> ...]] [--simulator <type>/<lens> [--dataset-types ...]] [--apply]:

• scaffold Science/<category>/<key>/ from a project template kept beside the conductor
  (AGENTS.md/CLAUDE.md skeleton, wiki/project/state.md, .gitignore, config/, scripts/,
  hpc/ with sync, sync.conf from the example, batch_cpu/, batch_gpu/), template derived
  from the current best project (active/euclid_detailed) with project specifics stripped;
• seed dataset/ by copying from Science/dataset/<sample> or running
  Science/simulators/<type>/<lens>.py --dataset-path dataset;
• add the projects.yaml row (ral_root mirroring the category) and cortex.py new the ledger,
  on a claude/birth-<key> branch — an add-only birth diff, which lands by itself under the
  existing ledger-merge rule;
• dry-run by default; --apply writes. Never creates a GitHub remote or touches RAL (prints the
  gh repo create / hpc/sync push-data-init lines for the human).

Out of scope

Moving trees on RAL; the human-readable dashboard work (draft/feature/cortex/output_summary.md).