[MaxGhenis]

Ruling d1724: fix U2's birth universe so §20 step 4 can run once more. Code and invented data only; no PSID record was read. career.py is byte-unchanged.

Why

The d1649 step-4 structure run (2026-10-19, 640adccf) stopped in cohort.derive_u2_births, before writing any output. career.derive_birth_years asserts that a birth year derived from a seed coordinate (age code 2–125) lies in [1889, 2016]. U2's universe takes every positive-weight, in-family person at any age, so a young child in a late wave, or an age code of 124–125 in 2013, reaches that assertion. The law also refuses, after its loop, every seed age code except 1–125 and 999. The assertion would have masked that refusal; the individual file's code 0 (Inap.) can reach it.

The rule (cohort._seed_for_the_law; spec §16d)

It applies only to persons whose birth year only the seed could give. Those are the persons the law, run without seed coordinates, leaves unresolved. For them:

• Refused seed code (0, or anything except 1–125 and 999): the seed becomes the earliest presence aged 2–125 whose implied year lies in 1889–2016 (re-seeded). If no such presence exists, the person is counted out under seed_age_0 (first code 0) or no_valid_seed_age.
• Out-of-support seed (age 2–125, implied year <1889 or >2016): the person is counted out under before_1889 or after_2016.

Everyone else keeps the law's precedence and record: a marriage-history or earnings year still wins, and codes 1 and 999 stay unresolved. A counted-out person never re-enters through the annuitant extension. That extension already seeds only presences the law accepts.

Disclosure: every row's structural counts carry n_seed_birth_year_counted_out (four keys) and n_seed_birth_year_reseeded. Each count decides nothing.

Reading choices (agreed with the coordinator; the widening approved by Max on 2026-10-20, noted on d1724):

• Only seed-dated persons are counted out, so no cohort member with a history or earnings year can be lost.
• Code 0 is re-seeded like any refused code instead of being counted out at once, because the loader documents 0 as Inap., not as an age. A cohort member with one bad presence keeps its seed. Max's approval named code 0 explicitly. It also named 999, but the law leaves 999 unresolved, so the rule keeps it on that path. That is narrower than what was described to Max, and §16d says so.

Invariants (tested)

• The universe and the four counted-out sets partition the presence set.
• Each person's outcome equals §16d's, computed from career run on that person alone. This is a Hypothesis property over every age code, bound and history/earnings mix, plus pinned examples.
• No counted-out person can be dated into 1946–55. A cohort member with one refused code keeps its seed and its observations.
• derive_u2_births completes on a fixture holding every case. The dry-run variant runs U1 and its income rows end to end.
• Base invented births are identical to master's; checked one-off by loading master's derive_u2_births from Git. Master raises on the counted-out fixture.
• Counting persons out changes nothing else: observations, dispositions, design frames, income rows and every other count stay as they were.
• The derivation is deterministic.

Spec

§3's lines are pinned to draft 3, so the "§3 note" is a new §16d. The d1090 record test now cuts §16c's record at §16d and requires a unique §16d heading. §16d is pinned line by line, with 21 text mutations. §17 points to §16d. §20 step 4 records that the d1649 pass stopped with no output and that d1724 authorizes one more pass. The §19 entry, the spec pin (u1_identity.json) and the exposure record are updated.

Checks

• U2 suite under an empty PSID root at the final head: 1,372 passed, 254 skipped. Estimates and Track M suites: passed.
• Code mutation check (EV/rulings-records-20261019-d1724/mutate.py): 24/24 killed. That includes the three mutants that survived the first review: re-seed support narrowed at both ends, narrowed at 2016 only, and the extension admitting code 0.
• Differential against master's derive_u2_births from Git: identical results on the base and all earlier invented variants. Master raises on both new variants. The first reviewer's 600-case random differential found no case where only this PR raises.
• Tiers by collection: master e0a731ae 9,916 / 4,253 / 1,423 / 520 / 308. This head (aa597e24): unit 9,958, artifact 4,253, integration_psid 1,450, 520, 308 (total 16,489).

Reviews (records: EV/rulings-records-20261019-d1724/)

1. 20261020-030251-d1724-review-552b, Opus 5.5 (no Codex lane free): APPROVE at 8a6a8192, with test gaps and record points. Addressed in 068ee2e2; one point declined (the version line is pinned by the d1090 check).
2. 20261020-045107-d1724-review-552-delta, GPT-6.1 Sol: APPROVE at 068ee2e2, one nit. Fixed in 88f3a172.
3. 20261020-051420-d1724-review-552-delta2 (one-line exposure-record fix): APPROVE at 88f3a172.
4. 20261020-114816-d1724-review-552-delta3b, Opus 5.5: REQUEST_CHANGES at 5dfc8413. It found that §16d left out the 999 divergence from the description Max approved, plus two wording points. Fixed in aa597e24.
5. 20261020-130902-d1724-review-552-delta4, GPT-6.1 Sol: APPROVE at aa597e24, with no findings.

axiom: n/a: U2 cohort-construction code and specification records, 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 20, 2026 8:21pm UTC

[MaxGhenis]

Merge audit (d1724). Merged with a merge commit as 1e739d328b73c99d259a72952b3452546721bfb8, pinned with --match-head-commit 7791ceaf7bffd13be41f7f10e5a6c22a3bf11ec2.

• Rulings:
  • Max's d1724 (2026-10-19) authorizes the fix.
  • Max approved its widening to seed codes the law refuses on 2026-10-20, including re-seeding code 0; this is noted on d1724.
  • The coordinating session (local_6085d52a-…) granted the tier slot.
• Reviews (records in EV/rulings-records-20261019-d1724/):
  • …-review-552b, Opus 5.5: APPROVE at 8a6a8192.
  • …-review-552-delta, GPT-6.1 Sol: APPROVE at 068ee2e2.
  • …-review-552-delta2: APPROVE at 88f3a172.
  • …-review-552-delta3b: REQUEST_CHANGES at 5dfc8413; fixed in aa597e24.
  • …-review-552-delta4, GPT-6.1 Sol: APPROVE at aa597e24, no findings.
  • …-review-552-merge: APPROVE at 7791ceaf. It confirmed the master merge brings in exactly Progress scorecard: add gate_epuf_fill and B2's failure #553's diff, byte for byte.
  • No review was outstanding with changes requested.
• Gates at 7791ceaf:
  • gh pr checks exited 0: lint, build, pytest (3.14), all four pytest shards and both graph-integration jobs passed.
  • mergeable was MERGEABLE (CLEAN), and the PR was not a draft.
  • Master was b7dae930, already merged into the branch.
• Checks:
  • 24/24 code mutants killed.
  • U2 suite under an empty PSID root: 1,372 passed.
  • Tiers by collection: unit 9,958, artifact 4,253, integration_psid 1,450, reproduction_legacy 520, oracle_policyengine 308 (total 16,489).
  • career.py is byte-unchanged.
• Data access: no PSID record was read, and no file on the restricted list was opened.
• Next: the coordinating session runs §20 step 4 once more under d1724 (structure, then component diagnostics), once master's Tests pass at the run head.