[MaxGhenis]

Records U2 specification §20 steps 2 and 3 as complete and §17 items 4–7 and 9 as far as they are done, each with its evidence. It follows the round-3 closeout of milestone 2 (#525, merged as f7a84d0b). INVENTED DATA ONLY: no real PSID record was read, and no restricted file was opened.

Review round 1 and what changed

The first independent review (Claude Opus 5.5, at e6a71a3c) returned REQUEST_CHANGES with one blocker. §20 step 2 called implementation done, but #525's report still listed one item as due before step 4. The frame digest hashes pandas' dtype names and CSV text, so the step-4 pass and the registered one-shot must run under the same pandas version and string-inference setting.

97eda904 closes that item in code, as the review preferred:

• cohort.frame_environment() returns {pandas_version, future_infer_string}.
• The pre-registration evidence record moves to schema v2 and freezes that environment with the digest. check_preregistration_evidence requires exactly a nonempty version string and a bool.
• loader.check_frame_environment refuses under any other environment:
  • a frozen load, before any PSID path is resolved;
  • a registered run, before the snapshot is built.
• 13 new tests, all on invented data, all failing on master's code:
  • the record carries the environment;
  • each malformed environment is refused with its own message;
  • a Hypothesis property: the registered check passes only under the frozen environment;
  • a frozen load refuses another version or setting without resolving a PSID path.

The review's six should-fix items and its informational wording are applied as it phrased them. Those cover the §17 review paragraph, step 3's citations, step 2's manifest counts and citation, item 4's citations, the SSI test's root patch, and the §19 blindness line. The milestone-2 report marks the pandas item closed.

What changes

The specification changes only after §16a, which the d1090 record tests require.

Place	Change
§20 step 2	Done: implementation (#487, #525 and this PR's frame environment), manifests (1,766 registry entries and 26 SSI and Census records, all RESOLVED and documentary), parameter capture (SSI, Track M, life tables pinned by hash)
§20 step 3	Done: independent mapping review (the milestone-1 adjudication, #486's verified corrections, and an independent review of every later registry change, each ending in APPROVE), invented tests, historical isolation, the exact U1 differential
§17 item 4	Done on the documents and invented data; real-record counts are the structural pass's
§17 item 5	Code pin checked; §15's pin is open until §20 step 5
§17 items 6 and 7	Done (the differential is evidence at f7a84d0b and runs again at ratification)
§17 item 9	Partly checked; the remaining equality checks wait for §20 step 5, which also empties §15's blocked_by
§17 review paragraph, §19	Updated; the §19 entry gains the blindness line and its exposure pointer

Other files:

• u1_identity.json: re-pins the spec SHA-256.
• D1090_FOLLOWERS: moves with item 9 and step 2.
• test_u2_parameters.py::test_a_changed_ssi_capture_refuses: refuses a copy of the capture with equal JSON content and one added space. It fails with either pin check removed (DID NOT RAISE).
• docs/design/u2_m2_u1_differential_f7a84d0.json: the differential at U2 M2 round-3 informational fixes: bind a registered run to its evidence record and rows #525's merge, equal in every check.

Tests

All runs used an audit hook refusing any PSID, evidence-directory or REFS access, with HOME and the PSID directory empty. The hook refused nothing.

• tests/track_u2 plus the U2 data, birth-evidence and legacy-manifest files: 1,345 passed, 254 skipped.
• Tiers, recounted by collection: unit 9,900 (+13), artifact 4,245 (+1), total 16,396. The manifest check passes.
• Black (26.5.1 and latest, -l 79) and ruff are clean.

Not in scope

• The structural pass (§20 step 4) needs Max's authorization, which the coordinator queues.
• §20 step 1's pinned clause "Source-verification blockers remain for steps 2–5" predates Apply Max's d637 ruling (amendments 1a-1c and 2-6) to the U2 source registries #540. It is pinned in D1090_CLAUSES, so a later revision should update it together with that test.

🤖 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 16, 2026 10:58pm UTC

[MaxGhenis]

Merge audit: merged as 2ef10532 (a merge commit) at head 0dd22555, 2026-10-18T16:11Z.

Landing slot. The "Microcosm dynamics" coordinator assigned this PR the tier-count slot after #542 landed. Master was 7b4eec2d.

Merge with master (0dd22555, master 7b4eec2d, which includes #542 and #545).

• The conflicts were the specification's SHA-256 pin, recomputed from the merged specification (77f30781…), and the two tier files.

• The tier files were recounted by collection on the merged tree:

  Tier	Count	Change from master
  unit	9,916	+14
  artifact	4,247	+3
  integration_psid	1,423
  reproduction_legacy	520
  oracle_policyengine	308
  Total	16,414	equals the full collection

• Merge delta (git patch-id --stable, excluding the conflicted files and the specification):

  • Each side equals its own changes since f7a84d0b.
  • The specification's master side is exactly Test a frozen U2 cited source at another hash; name the tier files in the exposure log #542's §19 line, and its branch side exactly this PR's §17, §19 and §20 edits.
  • The delta review confirmed both.

Independent reviews. All four were Claude Opus 5.5 reviewers with a shell, in a standalone clone with the PSID guard; the Codex lanes were held.

Run	Range	Verdict	Outcome
20261016-140335-pr544-review	e6a71a3c	REQUEST_CHANGES	One blocker: the pandas environment was still open before step 4. Six should-fix items.
20261016-152540-pr544-review-r2	97eda904	APPROVE	The blocker was closed in code: the evidence record (schema v2) freezes the frame environment, and frozen loads and registered runs refuse any other. Every should-fix item was applied.
20261016-185803-pr544-delta-review	97eda904..0dd22555	APPROVE	Nothing blocking and nothing to fix; three optional notes.

The reviews are in ~/reviews/pr525-closeout-20261015/out/.

Evidence at 0dd22555, under an audit hook refusing any PSID, evidence-directory or REFS access; the hook refused nothing:

• the d1090 record tests and every tests/track_u2 file plus the U2 data and birth-evidence files: 1,338 passed, 254 skipped;
• the legacy manifest: 11 passed;
• the U1 differential against 9cee2423: equal in every check, the dry run included;
• Black (26.5.1 and latest) and ruff: clean.

CI at 0dd22555: gh pr checks 544 exited 0 with every check passing; mergeable was MERGEABLE.

What this records:

• §20 steps 2 and 3 are done.
• §17 items 4, 6 and 7 are done, item 9 is partly checked, and item 5's code pin is checked.
• The last implementation item before step 4, the frame environment, is closed in code.

What remains:

• d1242 (B2 in sources.PERSON_COUNTED_EXCLUSIONS) follows as its own small PR.
• The step-4 pass (d1649) runs in a dedicated session, once, after both PRs have landed.

Optional notes from the delta review, left as follow-ups:

• The binding property test checks only the "changes" direction; the preflight tests cover the other.
• A docstring sentence order in scripts/track_u2_structure.py.
• _environment() could run before the load.

INVENTED DATA ONLY: no real PSID record was read, and no restricted file was opened.