[MaxGhenis]

Brings paper/paper.qmd in line with three results it still described as planned or absent. Facts are the same as #550's ledger update, each checked against code or committed artifacts.

B2 failed (#42 comment 6252752628; runs/track_b_b2_score.json, outcome: FAIL).

• The result goes in the summary (one sentence), the forward-earnings test sentence, the coverage table and the "Forward earnings in the projection" paragraph, which moves from future to past tense.
  • The deciding cell is the prime-age (25–44) zero-earnings rate: about 35% above the PSID's on a log-ratio tolerance of 0.202, about 22%. It misses on all five refit seeds, and only prime-age p90 passes every check.
• The consequences, from the design and the attempt ledger:
  • no rerun, new seeds or tolerance relief;
  • a successor whose changes must rest on evidence other than B2's result, registered before it runs;
  • two failed successors pause the mechanism.

gate_epuf_fill scored (#537; runs/epuf_fill_gate_test_v1.json). The Earnings Public-Use File paragraph no longer says we haven't scored anything against the file. It now describes the test from harness/epuf_fill_gate.py and states its results:

• How the test works: it holds back 20% of the file's people (878,481). It masks the odd years 1997–2005 and the years before max(1968, birth year + 22), fills them, and compares the fills with the true records.
• What it found: the blind tests' fill rules fail it, 100/183 and 131/136. The learned pre-career fill passes, 0/136, and the learned odd-year fill is adopted uncertified, 8/183. Neither is used by a blind test.
• Table rows: two table rows note that both current rules fail it.

Payment rules.

• The earnings-test paragraph now describes the built R2 part (a) ledger and grace year and the R3 modules (Track B R2 (part a): the retirement earnings test ledger #528, Track B R3: adjustment of the reduction factor and automatic recomputation #529). It notes that deduction before reduction is refused, and that the exercise-3 hypothesis test also needs forward earnings past 2010.
• The oracle paragraph adds the statutory delayed-credit window (Add the statutory delayed-retirement credit beside the sealed oracle; erratum for #494 #532) and the d1408 erratum diagnostic's finding (d1408: delayed-credit erratum diagnostic (counts and changes only) #538): no published value changes, and the first estimates are bounded, not recomputed.

tests/test_paper_text.py, test_anchor_provenance_erratum.py and test_track_b_b0_1_audit_record.py pass (38). The microcosm.institute paper bundle needs a rebuild to publish this. That rebuild must also include #536, which is still owed.

🤖 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 19, 2026 4:46pm UTC

[MaxGhenis]

Merged at reviewed head 38cf362 (merge commit e0a731a) by the dynamics coordinator. Gates: gh pr checks exit 0, MERGEABLE, not a draft, no change requests. Paper text only, so tier counts are unaffected. tests/test_paper_text.py passes.

Independent review ran three rounds, each on a standalone clone with a shell:

• Round 1 (GPT-6.1 Sol, 20261018-145018-md551-review): requested changes on eight points, each fixed in 8189111:
  • B2's seeds are household splits, not refits;
  • its 2012/2014 earnings come from the 2013/2015 interviews;
  • the passing statistic is the p90 of positive earnings;
  • the earnings-test refusal concerns the family maximum;
  • the limits section's EPUF sentences;
  • the summary's framing of the failure.
• Round 2 (Opus 5.5, 20261018-150125-md551-review-r2): APPROVE, with three optional points, applied in 38cf362: B2 in the abstract, and income years in place of "waves".
• Round 3 (Opus 5.5, 20261019-124532-md551-review-r3): APPROVE at 38cf362. The abstract's YAML parses and renders.

Publishing on microcosm.institute needs a paper-bundle rebuild, which must also carry #536.