[Deicyde]

Two mismatches between site/ops/brain_release_authority.py and what the Brain build actually reads.

1. The reducer's bot/state inputs were release-neutral

brain/authority/reducer-inputs-v1.json declares these as curated_git_input, read by brain/build_cells.py:

• bot/state/brain_queue.json
• bot/state/cut_log.json
• bot/state/pool_candidates.json (gitignored today)
• bot/state/recycle_queue.json
• bot/state/seed_queue.json

They sit under bot/, a neutral directory. So a commit that changed one let main move past a release built from the old bytes. That is the unsafe direction: the promoter and bundle freezer would accept a release that doesn't match main's reducer inputs.

They are now in AFFECTING_EXCEPTIONS. The rest of bot/state/ (bot_state.json, batchN_approved.json, …) stays neutral; it is bot operating state, not a reducer input.

2. The nightly's check is a different set, not a mirror

The comment on AFFECTING_EXCEPTIONS said that list must agree with brain-nightly.sh's uncommitted-source check. The two sets differ in both directions:

• The nightly watches more files than the policy treats as affecting. Among tracked files it refuses local edits to 54 files under wiki/test/, 4 under wiki/e2e/, wiki/README.md, two test_*.py files under wiki/scripts/, and site/ops/brain-canary.py. None of them reach release bytes, so refusing them is harmless. The canary entry is a leftover: the nightly stopped running brain-canary.py in afdf3a67, and the promoter and .github/workflows/brain-canary.yml run it now. The repo's own tests already treat the canary both ways: test_brain_nightly.py pins it in the nightly list, and test_brain_promote_release.py uses it as a neutral change.
• The nightly watches no curated inputs at all, neither bot/state/ nor catalog/data/. It expands $REDUCER_SCOPE to the inventory's scope code files only.

The comment now describes that instead of claiming agreement.

Not changed here, and worth a decision: the nightly runs brain/build_cells.py against its working tree. A local edit to a curated input on the nightly host would therefore be frozen into a release whose authority commit doesn't hold those bytes. Running python3 bot/brain_queue.py, which bot/README.md documents, is one way that edit happens. Closing that means adding the curated inputs to the nightly's source check. That is a brain-nightly.sh change, which is itself release-affecting, so I left it out.

Tests (site/ops/test_brain_release_authority.py)

• test_reducer_code_and_curated_inputs_are_release_affecting: across all four brain/authority/reducer-inputs-v*.json inventories, every reducer scope file, every curated_git_input path and every input rooted in the repository ("root": "repo") must be release-affecting. It fails on main for the five bot/state files.
• test_the_canary_is_the_only_nightly_watched_file_left_neutral: among the files the nightly names explicitly, the canary is the only one the policy calls neutral. It doesn't look inside watched directories like wiki/.
• New examples: bot/state/bot_state.json and batch7_approved.json neutral, bot/state/cut_log.json and seed_queue.json affecting.
• The nightly pathspec parser is shared as _nightly_watched() with the existing NEUTRAL_FILES test.

Docs

• docs/BRAIN-RELEASE-RUNBOOK.md, the operator runbook, now excepts the five bot/state/ files the reducer reads from its neutral list.
• bot/README.md's fold-back instructions note that folding any of the four tracked Brain-build files from bot-state into main is release-affecting, so the next promotion needs a new release.
• docs/BRAIN-ACTIVATION-2026-10-04.md is a record whose values can be checked at 37b7036d, so its list stays as it was. One sentence notes Release policy: AGENTS.md and CONTRIBUTING.md are release-neutral #68's change and this one.

Effect

• The module and its test are in site/ops/, so this PR is itself release-neutral.
• Tightening the policy affects existing bundles. The bundle freezer and validator re-check recorded neutral_changes against the current policy (brain_activation_bundle.py:719,1698), so a bundle that recorded one of the five files as neutral now fails validation. That is the intended result.
• POLICY stays wikilean.release-neutral-paths/v1, as it did in Release policy: AGENTS.md and CONTRIBUTING.md are release-neutral #68. That ID now names three rule sets (Releases survive release-neutral merges to main; pin-only edge diffs aggregate (semantic-diff/v3) #45, Release policy: AGENTS.md and CONTRIBUTING.md are release-neutral #68, this PR), and the promoter records it in intents, so a durable record can't say which rules it was checked under. Bumping it is a separate call.

Testing

CI runs the updated test file through scripts/ci-python.sh; the first push's run passed all 10 tests. I didn't run it locally because machine load was above the agreed limit. I checked the classifications directly: bot/state/cut_log.json is affecting, while bot/state/bot_state.json, bot/poll.py and site/ops/brain-canary.py are neutral.

[Deicyde]

Recommendation: fix (comment and doc wording only), then merge. An independent adversarial review of feed0d06 found the policy change and its tests correct. CI is green (10 policy tests). The PR is release-neutral, merges cleanly with #73, #74 and #78 in either order, and changes nothing operationally today: git diff --name-only 2eb31baa a8bb2882 -- bot/state is empty, so no recorded bundle lists these files.

• Fix: "the canary it runs" is false.
  • It appears in site/ops/brain_release_authority.py:45 and in the canary test's comment (site/ops/test_brain_release_authority.py, about line 110).
  • brain-nightly.sh mentions brain-canary.py only in its git status pathspec (line 833). afdf3a67 removed run_release_canary in September.
  • The promoter (brain_promote_release.py:3146) and .github/workflows/brain-canary.yml run the canary now. The watch entry is a leftover.
• Fix: tighten the same comment.
  • "watches code, not curated inputs, so it also refuses…" isn't a causal link.
  • "READMEs" is just wiki/README.md.
  • The check also watches brain/authority/schemas and reducer-inputs-v1.json, so "code" is loose.
• Fix: the docs' bot/state/*.json reads as the whole directory.
  • Where: docs/BRAIN-RELEASE-RUNBOOK.md:293 and docs/BRAIN-ACTIVATION-2026-10-04.md:133.
  • Only the five reducer inputs are affecting; bot_state.json, batchN_approved.json and the rest stay neutral. Say "the five bot/state/ files the reducer reads".
• Worth noting, not blocking:
  • pool_candidates.json is gitignored. Listing it only keeps the inventory test free of special cases. Its real exposure is the nightly host's working tree, and a future git status-based nightly check wouldn't see an ignored file without --ignored or an explicit check.
  • The test checks only scope and curated_git_input paths. A future repo-root immutable_source_object under a neutral directory would slip through. None exists today.
  • Folding bot state into main now forces a new release. That applies to any fold from bot-state that touches cut_log, seed_queue, recycle_queue or brain_queue. This is intended, but neither the runbook nor bot/README.md says so.
  • Pre-existing: .agents/ is not neutral while .claude/ is.

[Deicyde]

Follow-up: fixes applied in f3477291. Recommendation: merge. CI is green (run 39077219274; the 10 policy tests pass).

• Canary wording: the AFFECTING_EXCEPTIONS comment and the canary test's comment no longer say the nightly runs the canary. They say the nightly still names brain-canary.py but stopped running it in afdf3a67, and that the promoter runs it now. The loose "watches code, so…" causality is gone.
• Docs wording: docs/BRAIN-RELEASE-RUNBOOK.md and docs/BRAIN-ACTIVATION-2026-10-04.md say "the five bot/state/ files the reducer reads" instead of bot/state/*.json.
• Wider test: the inventory test also checks every input that v2 and later inventories root in the repository ("root": "repo"), not only curated_git_input. That is 47 paths across the four inventories, none neutral.
• Fold-back note: bot/README.md's fold-back steps say that folding any of the four tracked Brain-build files from bot-state into main is release-affecting, so the next promotion needs a new release.

Still open, as decisions rather than fixes:

• Nightly source check: should it refuse local edits to curated inputs? That needs --ignored or an explicit check for pool_candidates.json.
• POLICY ID: whether to bump it.
• .agents/: it is not neutral while .claude/ is.

[Deicyde]

Re-review on main (eaa5950b): fine. The five exceptions are exactly what brain/build_cells.py:101-105 reads, and removing any one fails the inventory test (checked with mutants). With #83 on top, the module comment, the tests and the nightly agree. Classifying every one of main's 18,659 tracked paths turned up no neutral path that reaches release, public-asset or Worker bytes. The inventory test's 47-path set has no neutral member.

Optional test hardening:

• test_every_file_the_nightly_names_is_release_affecting can't fail for directory entries such as site/assets: is_release_neutral() is never true for a bare directory string.
• The inventory test skips path_pattern inputs (site/annotations/*.json, catalog/data/user_repos/*.jsonl) and v1 entries that are neither curated nor root: repo. A mutant making catalog/data/user_repos neutral passes. Nothing is misclassified today.

Stale docs:

• bot/.gitignore:6-7 says pool_candidates.json "only feeds the /queue publish". build_cells.py reads it.
• docs/BRAIN-RELEASE-RUNBOOK.md:291 calls test_*.py neutral without the site/test_frontier_page.py exception.

Outside this PR, and the most important item:

• Attest the non-Brain public assets for Brain release B #82 attests atlas assets that main lacks. It attests assets/atlas.css and assets/atlas.js in wiki/public-asset-source-attestation.json, but main has no source for them. Formalization atlas: a read-only data plane for where WikiProject Mathematics is formalized #85 is still open, and build-public.ts doesn't copy them. A public-baseline freeze at main therefore fails, or passes only with a tree built outside main. Settle that before release B.
• The nightly's source check misses some reducer imports, even though the policy correctly treats them as affecting: install_wikidata_observation, wikidata_observation, catalog/huggingface_download.py and brain/tools/execution_environment.py.