[MaxGhenis]

This adds U2's pre-registration evidence record at docs/design/u2_preregistration_evidence.json, the path #556 made loader.PREREGISTRATION_EVIDENCE_PATH. From here on the record's presence is the intended state. Four checks asserted that the record was absent, so they now treat it as present. Every refusal branch is still tested: absence is forced through an explicit missing path, and the record is never removed.

The record (35954275)

• A byte-for-byte copy of the record that the §20 step-4 pass wrote under d1724 at 1e739d328b73c99d259a72952b3452546721bfb8 (master Tests run 39502262845). It holds hashes only.
• SHA-256 7bc5c3c26ff014271061a8228068c6e71103430ec299048462570476a510b845, 13,946 bytes. Checked equal at the source, at the destination (cmp byte-identical) and in git show HEAD:docs/design/u2_preregistration_evidence.json. It equals the step-4 pass's MANIFEST pin and loader.PREREGISTRATION_EVIDENCE_SHA256.
• Value-free scan: CLEAN. The scan report's SHA-256 is d1f1b1bc3ca1e1f4ee8a899c2ddede2858f2cf3b54b6138179127e0e47872dd0.

Tests treat the record as present (2faf0356)

1. tests/track_u2/test_u2_loader.py: test_the_committed_evidence_record_is_absent_so_registration_refuses is replaced in place by test_the_committed_evidence_record_is_the_pinned_one. It asserts:
   • the file exists;
   • its SHA-256 equals loader.PREREGISTRATION_EVIDENCE_SHA256 and the literal pin;
   • read_preregistration_evidence() accepts it;
   • read_preregistration_evidence(tmp_path / "absent.json") still refuses with "no pre-registration";
   • a copy one hex digit away refuses as "not the committed record". That copy is still well-formed (check_preregistration_evidence accepts it), so only the pin refuses it.
2. test_a_frozen_load_needs_the_committed_record keeps its "committed record differs" case.
   • The "no record" case now patches loader.read_preregistration_evidence with a wrapper that reads a missing path. Patching PREREGISTRATION_EVIDENCE_PATH would not work, because the default argument is bound at definition.
   • It still refuses before any PSID path is resolved or any file opened.
3. tests/track_u2/test_u2_scripts.py::test_preflight_refuses_while_the_evidence_record_is_absent uses the same wrapper for the registered preflight.
4. scripts/track_u2_dry_run.py: registered_preflight_without_evidence now reads an explicit path inside a fresh temporary directory, and checks that the path does not exist first. The check raises rather than asserts, so it holds under python -O. The output keys are unchanged, and test_dry_run_records_every_branch_and_refusal passes unchanged.

Not touched: docs/design/boomers2004_1946_55_comparison.md (the registration binds its SHA-256), loader.py and scripts/run_track_u2_registered.py. No test is added or removed. .test_durations is not edited here, so it still lists the replaced test's old name.

Test results

All runs used an empty PSID root (POPULACE_DYNAMICS_PSID_DIR and HOME pointed at empty directories), -n 4, Python 3.14.7 and a .[dev] install (pandas 3.1.0, numpy 2.5.4).

• At 2faf0356: tests/track_u2 plus tests/estimates/test_birth_evidence_artifact.py (which lists the dry-run script): 1,432 passed, 254 skipped, 0 failed.
• tests/track_u2 compared with 0e462efa (before the record existed, 1,419 passed, 254 skipped): per-test outcomes are identical, except that the replaced test passes under its new name.
• Tier counts (pytest --collect-only -q -m <tier>), unchanged before and after the edits and equal to tests/tier_counts.json:
  • unit 18,125
  • artifact 4,261
  • integration_psid 1,450
  • reproduction_legacy 520
  • oracle_policyengine 308
  • 24,664 in total
• Black and Ruff are clean on the three changed Python files.

Exact U1 differential at 2faf0356 (c9c0f4a2)

docs/design/u2_m2_u1_differential_2faf035.json (SHA-256 4d19093b56fe52555388cff7ee721b80e8302c7a859bd6a18d077e100229c891) is the output of scripts/u2_u1_differential.py, run the way #556 ran it and in the frame environment the evidence record pins.

Setup

• base 9cee2423f048683e15838aa5fee330afd072f85f and candidate 2faf03565cb55ab56256317aebefab2a2ce82efd, as clean detached checkouts;
• Python 3.14.7, pandas 3.0.6 and numpy 2.5.4 (the step-4 pass's versions), with future.infer_string True;
• seed 20260924 and an empty PSID root.

The report's environment block records these settings. It was read from the same interpreter after the run. With the block removed, the file is byte-identical to the harness's output.

Results

• all_equal true and all_checks_passed true.
• Refusal expectations met on both sides.
• loader_file_accesses_empty true, and dry_run_skipped false.
• The contract-check digest (49e47726…) equals the one in u2_m2_u1_differential_14d9576.json, and in every earlier record.
• The dry-run digests are equal between base and candidate. They differ from earlier records' digests, as those differ from one another, because the harness's temporary output directory is part of the compared run.command.

Under pandas 3.1.0, which a fresh .[dev] install and CI now resolve, the results are the same. That report was committed at this path in a0fd18cc (SHA-256 d85499e5…, numpy 2.5.4). It differs from the 3.0.6 report only in the interpreter path, the temporary directory in run.command, and the dry-run digests that include that directory.

Carry-forward for step 9: the registered run's pandas

The registered one-shot must run under pandas==3.0.6 with future.infer_string True: the frame_environment the evidence record pins.

• Why it matters: a fresh .[dev] install now resolves pandas 3.1.0.
• What happens under another pandas: loader.load_u2_inputs(evidence=…) calls check_frame_environment. That check compares {pandas_version, future_infer_string} with the record and raises U2FrozenIdentityMismatch before any PSID path is resolved. scripts/run_track_u2_registered.py calls that loader. runner.run_track_u2 checks the same environment again and raises U2RunError.
• numpy: the record does not pin it; the step-4 pass ran numpy 2.5.4.
• Install used here for the differential: uv pip install -e ".[dev]" "pandas==3.0.6" "numpy==2.5.4" under Python 3.14.7.
• The test suite: CI's pandas 3.1.0 is fine for it.

Review and CI at c9c0f4a2

• Independent review: APPROVE, no findings. Done by GPT-6.1 Sol (Subfleet job 20261024-121913-md577-review-c9c0f4a) in its own clone at this head, with an empty PSID root. The reviewer ran every check itself:
  • Scope: five files and four commits; the protected files and tests/tier_counts.json are unchanged.
  • Mutations: pointing either wrapper at the committed path makes its test fail. Under python -O the dry-run helper still raises before the read.
  • Dry-run output: the key paths of result.json equal those at 0e462efa (29,004 paths).
  • Tests: 1,432 passed, 254 skipped, 0 failed. The five tier counts are unchanged, and the only node-ID change is the renamed test.
  • Lint: Black and Ruff are clean.
  • Differential: the reviewer reran it independently under pandas 3.0.6. It found no field differences after normalizing paths and the temp directory, and it reproduced both committed dry-run digests by changing only the temp-directory string in run.command.
  • Frame-environment claim: checked against the code and by direct probes.
• CI: Tests run 40110842940 succeeded at c9c0f4a2, and every check passes.

axiom: n/a: committed evidence record (hashes) and test updates, no policy rule

🤖 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 24, 2026 4:18pm UTC

[MaxGhenis]

Closing and reopening only to rerun CI on current master (4e20c6e4, after #582). The head is unchanged.

[MaxGhenis]

The dynamics coordinator merged this at its reviewed head c9c0f4a20a1d37526acca7346851747e10e611eb (merge commit 9967dfd6eb26860f562696e82b1ab03f305a1c81), with --merge --match-head-commit.

Gates.

• gh pr checks exited 0 on the latest run: four pytest shards, pytest (3.14), graph-integration (3.13 and 3.14), lint, build and Vercel, with deploy skipped. The earlier green run predated Track B: held-out data registry for every Track B and open-population gate #560's history test, so the PR was closed and reopened at the same head to rerun CI on current master after Registry: compare a pull request's own commits from its fork point #582.
• The PR was MERGEABLE and CLEAN, not a draft, and no review requested changes.
• tests/tier_counts.json is unchanged. The PR adds no test; it edits two test modules so the four absence tests force absence through explicit paths. Master's later merges (OP0.1: source, exposure, access and power audit for the demographic gates #567, Parity ledger: G v2, S1, the three approved designs, U2's ratification and the known gaps #569) touch none of its files.

Independent review. GPT-6.1 Sol (20261024-121913-md577-review-c9c0f4a, the U2 lane's job): APPROVE at c9c0f4a2, no findings. It confirmed:

• the scope is exactly five files and four commits over 0e462efa;
• the protected files and the tier counts are unchanged;
• the dry-run output at HEAD and at 0e462efa has matching JSON key paths.

Next. U2 §20 step 8, the #42 registration, is Max's decision. The registration binding is computed at this merge commit before the comment goes to him.