Track B: keep B2's outcome files out of every blinded review root

[MaxGhenis]

What

This PR keeps B2's outcome files out of every blinded review root.

B2's run record (#547) and the cumulative attempt ledger (#548) hold reference-year 2012 and 2014 floor values, truth moments and B2's candidate outcome. No exposure-inventory scan lists them, so a root built from master would include them.

• The record. docs/design/track_b_b2_outcome_exclusions.json is a standing exclusion record, in the inventories' q6_exclusions form. It names five files:
  • runs/track_b_b2_floor.json
  • runs/track_b_b2_floor_v1_stopped.json
  • runs/track_b_b2_score.json
  • runs/track_b_b2_score.attempt.jsonl
  • docs/design/track_b_attempt_ledger.md
• The tool. scripts/track_b_b0_1_review_root.py gets a STANDING_EXCLUSIONS tuple and standing_inventories(). build() now unions the standing records with whatever --inventory files a builder names, so the guard enforces the exclusion even when a builder forgets it. It reads every exclusion record before it writes anything. A missing standing record therefore refuses the build, rather than dropping silently, and leaves no partial root.
• What stays the same. The B0.1 exposure inventories are scanner outputs whose hashes the record tests pin, so they are unchanged.

The dynamics coordinator added the same paths to the evidence folder's RESTRICTED-FILES.md for blind roles. This PR makes the root builder enforce it.

Invariants

• For any set of named inventories, the denied paths are a superset of the standing record's paths, and exactly the union of the named and standing paths, sorted with no repeats. This is a Hypothesis test with 100 examples.
• A build that names no inventory still removes all five files, and its record lists them as denied and removed. This is tested end-to-end on an invented archive.
• A named inventory that repeats a standing path and names a path the archive lacks gives a record that denies the sorted union. Only the files the archive held are removed.
• A missing standing record raises FileNotFoundError before any output exists.

Tests

• New: 6 tests, in tests/test_track_b_b2_outcome_exclusions.py. They are artifact tier, since the file names runs/…json paths, so artifact goes from 4,247 to 4,253.
• Local runs. The new file, test_track_b_b0_1_review_root.py, test_track_b_b0_1_addendum_record.py, test_track_b_b0_1_revision1_tools.py and test_track_b_b0_1_audit_record.py gave 310 passed. Black 26.5.1 -l 79 and ruff are clean.
• Mutation checks: 4 of 4 killed:
  • build() ignores the standing records;
  • a missing record is skipped;
  • the record drops runs/track_b_b2_score.json;
  • the standing records are read after the output is created.

Review round 1 (Astra, job 20261018-142942-trackb-pr549-r1): APPROVE, with no blocking finding. It reran the three original mutants, all killed, and got 307 passed. It skipped one test that opens restricted files. Its two nonblocking suggestions are both taken in 00786d79: read the records before writing, and test the union inside build().

axiom: n/a (review tooling)

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
social-security-model	Ready	Preview	Oct 18, 2026 6:51pm UTC

[MaxGhenis]

Merged at head 00786d79 as merge commit a61ce746 (2026-10-19 16:45Z).

Gates

• gh pr checks exited 0 on the latest run, and the PR was MERGEABLE and not a draft, with no changes requested.
• The dynamics coordinator gave the tier slot. Master tiers are now unit 9,916 / artifact 4,253 / integration_psid 1,423 / reproduction_legacy 520 / oracle_policyengine 308.

Reviews

• Round 1 (Astra, job 20261018-142942-trackb-pr549-r1): APPROVE, with no blocking findings. It reran the three mutants, all killed, and got 307 passed. It skipped one test that opens restricted files. Its two nonblocking suggestions are taken in 00786d79.
• Round 2 (Astra, job 20261018-145143-trackb-pr549-r2): APPROVE, with no findings. It reran the read-order mutant, which was killed, and got 147 passed and 6 artifact tests collected.

From this merge on, every blinded review root built by scripts/track_b_b0_1_review_root.py excludes B2's run record and the attempt ledger, whatever inventories its builder names.