[MaxGhenis]

S0.1 of the mortality, groups and financing lane: the source, exposure, access and power audit for the gates SM2, SM3, SM4, ED1 and ED2. It is validation-only. It freezes what each gate's later registered test will use, and computes power on invented data. It estimates nothing from real data.

The coordinator lands this PR at its tier-count slot.

Sources:

• The lane's design: docs/design/mortality_groups_financing.md, approved under d1922 (Design: mortality, groups and financing lane (priced milestones, approved under d1922) #558, at 2c19f5b6). S0.1 is in its now-set, after SM1 (SM1: differential mortality by lifetime earnings, education and race (normalized within age-sex-year cells) #570) on lane M.
• The builder brief: s01-brief-20261024.md, in the lane's evidence folder.
• Out of scope: the group gates (GR1 to GR4). Their audit is briefed separately.

What it adds

• docs/design/s0_1_source_audit.md: the frozen record, one section per gate, in the order SM3, SM2, SM4, ED1, ED2. For each gate it records:

  • the sources, with their manifest hashes;
  • the population and estimands;
  • the exposure history;
  • the holdout partition and prediction origin;
  • the critical cells and tolerances;
  • the uncertainty treatment;
  • the pass conjunction and power procedure;
  • the suppression threshold;
  • the named conditions.

  It marks each item frozen, proposed or open. It also has a code-input metadata table (each repository input's path, commit, SHA-256, verification date and reliance). It ends with the input for Q7, what SM3's real-data step would need, and makes no recommendation on authorization.

• scripts/s0_1_power.py and its committed results, docs/design/s0_1_power_results.json. The script simulates SM3's and ED2's registered tests on invented populations, over a declared grid of assumed counts, and reports power at the 0.90 criterion. Its threshold helpers return the minimal sufficient count or raise a named refusal.

• tests/s0_1/: 160 example and Hypothesis property tests on invented data.

• Housekeeping: the tier counts.

Invariants

Tested as properties (Hypothesis) and examples, on invented data:

• simulated type-I error under the null is within Monte Carlo error of the nominal level;
• power is nondecreasing in sample size and in effect size;
• fixed seeds are deterministic;
• Bonferroni bounds cover at least nominally across the registered family;
• threshold helpers: for every accepted input, the returned count meets the helper's own floating-point predicate and, above the smallest supported count, one fewer does not. The supported counts are the integers from one through MAX_SUPPORTED_COUNT. Every other input ends in a named refusal (ThresholdInputError, UnreachableThreshold or UnrepresentableThreshold, all ThresholdRefusal subclasses of ValueError), never None or another exception;
• failure preservation: an earlier FAIL survives any later inestimable cell, and inestimability never produces a PASS;
• malformed inputs are refused, and insufficient precision is inconclusive, never a pass.

Review r3 found two places where earlier code broke an invariant it stated. Round 3 fixed both, and round 4 fixed the two numerical gaps review r4 then found in the threshold helpers. Round 4 deliberately refuses some inputs round 3 returned a result for; the note lists those kinds. The recorded planning and validation counts are unchanged.

Checks

• On the host, at the committed tree, with an empty PSID root: tests/s0_1: 160 passed, under an audit hook that counted opens of the files the project bars from these runs. It counted none, and no populace_dynamics module was loaded. The full-collection tier check runs at the coordinator's slot.
• Values scans (validation-only jobs) after each round: scans r5, r6, r8 and r9 covered rounds 3, 4, 5 and 6, and each returned CLEAR. Each audited the fix job's every read from its event stream, and checked every changed line against the restricted sources. They found no line in a restricted class and no carried-over value.
• A first attempt at round 5 was set aside. Its values scan, r7, found no carried-over value but could not certify the job's reads, because the job had searched code directories more broadly than its brief allowed. The lane session restored the round-4 note and reran the round as fix r5b under stricter search rules. Scan r8 certified r5b's reads.
• Review r5 20261026-190834-mgf-s01-review-r5 (GPT-6.1 Sol): APPROVE, with no high or medium finding.
  • It proved sufficiency and strict minimality for the helpers and tried to break them on plateaus, underflow, adjacent target floats and the domain's edges. It found no counterexample, and no caller that swallows a refusal.
  • It reran the script without its anchor inputs; every non-anchor field matched the committed results exactly.
  • It checked that review r4's two medium findings and its qualification are resolved, and that no earlier fix regressed.
• Review r6 20261026-215157-mgf-s01-review-r6 (GPT-6.1 Sol), on round 5: APPROVE, with one low wording finding, fixed in round 6.
  • It checked the new support definition against the code, and probed the list of refused inputs on invented argument, type and boundary cases, and found no kind the note misses.
  • It confirmed that review r5's two findings are resolved and that no earlier fix regressed.
• Review r7 20261026-223940-mgf-s01-review-r7 (GPT-6.1 Sol), on round 6: APPROVE, with no findings.
  • It confirmed from the code and invented cases that both helpers can return one and never evaluate zero, and that both statements of the one-fewer property now except a threshold of one.
  • It replayed review r6's finding, checked every minimality statement in all nine of the PR's files, reran tests/s0_1 (all passed), and found no earlier fix narrowed or regressed.
• The committed blobs match the SHA-256 hashes scan r9 and review r7 covered.

Tiers (on master 6413d842):

Tier	Before	After
unit	19,504	19,664
Total	26,074	26,234

The other tiers keep master's counts. The 160 new unit tests are tests/s0_1: a named collection on the merged tree gives 160 unit tests and none in any other tier, and CI's full collection checks the total. The master merge took master's two tier files, and one later commit recounts them. This PR's seven other files keep the blobs review r7 covered at 781ba61f.

Provenance

• Built by Subfleet job 20261024-110212-mgf-s0-1-build.
• Revised in six fix rounds: 20261024-211800-mgf-s01-fix-r1, 20261025-015100-mgf-s01-fix-r2b, 20261025-114320-mgf-s01-fix-r3, 20261026-162925-mgf-s01-fix-r4, 20261026-203129-mgf-s01-fix-r5b and 20261026-220958-mgf-s01-fix-r6. The set-aside attempt was 20261026-194239-mgf-s01-fix-r5.
• Values-scanned after the build and after each round: 20261024-171232-mgf-s01-values-scan, then delta scans r2 to r9 (the last 20261026-222719-mgf-s01-values-scan-r9).
• Committed by the lane session (Claude Opus 5.5): the fix jobs left their trees uncommitted.
• Data: invented data only. PSID use follows the project's data-access record.

Review

• Rounds 1 and 2b returned REVISE; fix rounds r1 and r2b addressed them.
• Round 3 (20261025-105651-mgf-s01-review-r3) returned APPROVE with five low findings, fixed in round 3.
• Round 4 (20261026-131614-mgf-s01-review-r4) returned REVISE with two medium findings in the threshold helpers' numerical handling: a reachable finite target could return None, and some numerical endpoints raised raw exceptions. Round 4 fixed both.
• Round 5 (20261026-190834-mgf-s01-review-r5) returned APPROVE with two low documentation findings: the support definition omitted the cap, and the preservation claim was unqualified. Round 5 fixed both.
• Round 6 (20261026-215157-mgf-s01-review-r6) returned APPROVE with one low finding: the note's statement that each reported threshold passes and one fewer fails needed the lower-end exception, since the smallest supported count is one and zero is unsupported. Round 6 fixed it.
• Round 7 (20261026-223940-mgf-s01-review-r7) returned APPROVE with no findings.

axiom: n/a: source audit and invented-data power calculations

🤖 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	Nov 1, 2026 6:18pm UTC

[MaxGhenis]

The Microcosm Dynamics coordinator merged this PR at 502915ea54dfad3cb21837fcdcfb90f238845505 with --merge --match-head-commit, as 1808285925a069143047ae7e7bb2412012db0ad7 (19:31Z, 2026-11-01).

• Review. Review r7 (GPT-6.1 Sol, job 20261026-223940-mgf-s01-review-r7) approved 781ba61f with no findings. The two later commits change no reviewed content:
  • 1f46d682 merges master 6413d842. The PR's seven non-tier files keep their 781ba61f blobs, and every other path equals master's.
  • 502915ea recounts: it touches only tests/README-tiers.md and tests/tier_counts.json, and adds the 160 unit tests in tests/s0_1 (unit 19,664, total 26,234).
• CI. gh pr checks 588 exited 0 at the head, 11 checks passing and deploy skipped: run 41335801487, with its four pytest shards and graph-integration 3.13 and 3.14. The full collection in the tier-policy test confirms the total.
• The slot now. Master is 18082859. Each remaining tier-slot PR merges it, or later, and recounts.