Register gate_epuf against SSA's Earnings Public-Use File (unlocked, report-only)

[MaxGhenis]

What this does

Registers gate_epuf: the rules for scoring the gate-1 generator's earnings against SSA's 2006 Earnings Public-Use File (EPUF), a public 1 percent sample of Social Security numbers with capped taxable earnings for 1951-2006. A gate is a pass-or-fail test whose rules and thresholds are fixed and published before the model is scored against it.

Outcome: the gate does not lock. As registered it has not been shown to catch anything gate 1 does not already catch, and it could not meet its own check on its bite. It gates nothing. Nothing here edits gates.yaml, a locked gate or a committed runs/*.json; the registration record is docs/design/gate_epuf_block_draft.yaml (locked: false, status: unlocked_report_only). No model has been scored.

Full record: docs/amendments/gate_epuf_registration_proposal.md. Review rounds, each by an independent Opus 5.5 lane: reviews/gate_epuf_round1_referee_20261002.md (AMEND: do not lock), reviews/gate_epuf_round2_verification_20261002.md (merge after listed fixes), reviews/gate_epuf_round3_rereview_20261002.md (APPROVE, five minor points), and reviews/gate_epuf_round4_confirmation_20261002.md, a confirmation of 95291e75 (APPROVE, three optional points; also posted below as a comment).

Why it does not lock

• The generator redraws earnings on held-out persons' observed PSID periods (even years 1998-2022, ages 25-59), so its overlap with EPUF's reliable years is 1998, 2000, 2002 and 2004. Nothing generates a career's earnings before 1998.
• On that overlap only two cells met the rules for gating: the 1998-2004 rank persistence of capped earnings, for men and for women.
• Those cells fail a persistence shortfall from the PSID of about 0.10 four times in five and 0.11 nine times in ten. The registration's own check asked them to catch a smaller shortfall (0.07 for men, 0.06 for women) nine times in ten, which the rules could not deliver. The pause that triggered was a flaw in the registration, not a surprise in the data.
• A perturbation drawing each person's early years from donors of both sexes raised men's persistence to about EPUF's level, which the cells accept. No perturbation was shown to pass gate 1 and fail these cells.
• With the PSID's distance from EPUF now public, candidate 11's verdict on these cells is largely predictable.

The round-1 referee (independent Opus 5.5) rejected loosening the check after the result. This PR takes that ruling. Section 12 of the proposal lists what a gate with bite would need: cells whose bridges no one has computed (which needs a generator of earlier years or careers), a bite dosed above the cell's own detection point, and a demonstrated catch beyond gate 1.

What EPUF does measure

On the generator's support (5,769 PSID persons; 1,311,282 EPUF persons born 1947-1973). Each figure is the unweighted mean of three birth-cohort bands.

	EPUF	PSID
Rank persistence 1998-2004, men	0.711	0.668
Rank persistence 1998-2004, women	0.672	0.640
Men's positive person-years at the taxable maximum	11.6%	17.8%
Men with a zero year in 1998-2002, given earnings in 2004	12.8%	9.7%

The persistence gap holds in five of the six sex-by-cohort cells. The artifact measures these gaps; it does not separate how much comes from reporting, from who the PSID samples, or from noncovered work.

On EPUF alone, rewriting careers by two of the career assembler's rules (nothing before 1968; odd years from 1997 filled from neighbours) lowers the median AIME of men born 1940-1944 by 10 percent and of men born 1930-1934 by 36 percent. A cohort born in 1946 or later loses no year to the first rule.

EPUF has 4,384,254 persons; the 4,348,254 in parts of SSA's article is a digit transposition.

What a run reports

populace_dynamics.harness.epuf_run.report_candidate reports every window cell for a candidate's 20 generated panels: the 20-seed estimate, its distance from EPUF, and that distance split into the candidate's distance from the PSID and the PSID's distance from EPUF. It returns no pass or fail. No run script calls it yet.

Invariants, and where they are tested

• The reader refuses bytes other than the pinned SHA-256, and reproduces SSA's published Table A1, Table 4, Charts 3-4 and the research note's Table 8 from the staged file (tests/data/test_epuf.py).
• The measurement operator keeps every value in [0, wage base], preserves positive and at-maximum status, and is weakly increasing in all 56 years (Hypothesis, tests/harness/test_epuf_operator.py).
• Unit-weight Spearman equals scipy's; integer weights equal repeated rows; cells ignore weight scale and row order; the career AIME equals ss.statutory_aime.aime (tests/harness/test_epuf_cells.py).
• The acceptance interval always contains EPUF and the PSID's position and never extends past EPUF by more than the tolerance; an eligible cell's interval lies inside the cap (tests/harness/test_epuf_gate.py).
• The two-term floor matches the 20-seed spread of a generator that draws from the true law, in simulation, and a holdout-only floor is too narrow (test_floor_prices_a_faithful_generator).
• A report's gap equals its source term plus its model term in every cell, and a report carries no pass field (tests/test_epuf_gate_floor_builder.py).
• The floor artifact, the block and the supplement recompute from stored replicates; the derivation files hash to what the artifact recorded; the first build's window results equal the rebuild's; the reporting path and supplement are pinned by literal SHA-256 (tests/test_gate_epuf_block_draft.py).

Commits

• eec910d6 rules first, pushed before any real-PSID value in EPUF units existed.
• 970a9db7 fixes the report-only career AIME (it now ranks every year after 1950, as the statute does).
• ce8d5000 floor artifact, block, results.
• 246eab02 excludes the new modules from the birth-evidence reducer's identity seal (the earlier CI failure).
• d606ddea referee round 1: ruling, supplement, frozen first build, corrected wording, paper sentence.
• 93848324 verification round 2: a reporting path with no verdict; wording.
• 95291e75 the approving re-review's five minor points.
• 56938042 the confirmation review's three optional points: an independent check of the pooled estimate (on seed values now skewed, since a median passed the old symmetric ones), verdicts kept apart from fixes in the block's review records, and one rewrapped line. The block now records round 4.

The five derivation files are unchanged since ce8d5000, so the floor artifact still binds to the code that built it.

Paper

paper/paper.qmd, section "The PSID's limits": the last sentences of the EPUF paragraph no longer say the comparison is undesigned and unregistered. They say its rules are registered for the generated earnings, that only two cells met the rules for gating and nothing was found that they would catch beyond gate 1, so the comparison is reported without a pass or fail, and that the model has not yet been scored. The site's paper bundle needs a rebuild after merge.

Tests

150 new tests (120 unit, 30 artifact); the EPUF reader's reproduction tests skip unless the pinned bytes are staged. tests/tier_counts.json and tests/README-tiers.md are updated.

🤖 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 4, 2026 3:24pm UTC

[MaxGhenis]

Confirmation review of the final head 95291e75 (independent Opus 5.5 lane, subfleet job 20261002-173551-epuf-gate-r4; read-only, no shell). Posted verbatim. Rounds 1-3 are committed under reviews/.

APPROVE

All five of round 3's points are fixed at 95291e75. The delta adds no false claim and nothing that reads as a lock or a certification. gates.yaml still has no EPUF entry, and the block still has locked: false and gated_cells: {}. I had no shell, so this rests on reading the files; I didn't run the tests.

#	Round-3 point	Status	Evidence
1	"only two cells had enough power" in the paper and the proposal's outcome	RESOLVED	Both now say "only two cells met the rules for gating" (paper/paper.qmd:366, proposal.md:36), the wording round 3 suggested. The sentence is now true and matches §7.2–7.3. §7.2 says r6.women.c1 is eligible but not gated, because the ladder gates cohort cells only when all six are eligible (proposal.md:374-377). §7.3 says "The ladder fell back to sex-level persistence and selected two cells" (:381). The rest of the paper sentence ("we found nothing they would catch that gate 1 does not… without a pass or fail") matches §7.4 (:430-431). No "enough power" or "had the power" is left in the paper, proposal, block or renderer.
2	report_candidate tests: source and model terms on nonzero bridges, and the NaN paths	RESOLVED	See the notes below the table.
3	Supplement tests could pass vacuously	RESOLVED	See the notes below the table.
4	§10 commit record and the checklist wording	RESOLVED	Item 4 now names d606ddea. A new item 5 names 93848324 (report_candidate, the round-2 report and its corrections) and mentions the final commit. The checklist says "a pinned reporting path" and adds the re-review line. "Scoring path" survives only in the block's round-2 record (block.yaml:206, renderer :73), where it correctly describes what was replaced.
5	Four wording fixes	RESOLVED	• Block ruling (rendered from the script, block.yaml:190-199): "0.067 for men and 0.057 for women". This matches §7.4 and the supplement's −0.0665/−0.0568. • §10 quotes "the house floor formula", the docstring's exact text (epuf_gate.py:179). • §7.4 says bd2c moves the cells "about 0.01", matching the table's +0.007/+0.010 (proposal.md:402). • §9 says tranche R "would report… no code computes its PSID side yet". career_cells is called only on EPUF data (build_epuf_gate_floors.py:295-296), so this is accurate.

Point 2 in detail. The new tests cannot pass vacuously:

• Real bridges, not the synthetic fixture. test_report_candidate_terms_against_the_committed_bridges feeds the committed artifact's cells into report_candidate. candidate_window_cells is stubbed out (monkeypatched); report_candidate calls it through the module, so the stub takes effect.
• Nonzero bridges. I counted the artifact's 41 bridges: 38 have |b| > 0.01, against the test's floor of 25. Only three are below 0.01 (0.0045, −0.0097, −0.0095).
• Source term. The test compares the code's psid − epuf on the cell's scale with the stored bridge_psid_minus_epuf. A flipped sign or a wrong scale would now fail.
• Model term. It is checked as estimate − transform(psid), which is computed separately from the code's gap − source.
• Coverage. The test asserts that the report's cell ids equal cell_ids(), so all 41 cells are checked. It also asserts that neither the report nor any cell carries a pass key.
• Defined values only. pytest.approx never treats NaN as equal to NaN, so the test cannot pass by producing NaN.
• NaN paths. test_report_candidate_propagates_undefined_values sets one seed of r6.men to NaN, which makes its estimate and gap NaN. It sets r6.women's psid_value to None, which leaves the gap defined but makes both split terms NaN (via _on_scale, epuf_run.py:102-103).

Point 3 in detail.

• set(bites[name]["cells"]) == set(artifact["registered"]) is added (test:161). It can't hold for empty sets: the detection-point loop indexes bites["bd1_persistence_loss_0.10"]["cells"]["r6.men"] and ["r6.women"] (:209), which would fail if those cells were missing.
• set(bites["detection_points"]) == {"r6.men", "r6.women"} is a literal (:192).
• The band years are now tied to COHORT_BANDS (:350-351), which defines c0 = 1947–1955, c1 = 1956–1964 and c2 = 1965–1973.

New findings. Nothing blocks merge. Three optional nits:

• The pooled estimate is checked against itself. Test 1 compares estimate with gate.pooled_estimate, the same function the code calls, so an error in that function would go unnoticed. Fix (optional): also assert row["estimate"] == pytest.approx(transform(cell_id, statistics.fmean(per_seed))).
• The round-3 verdict mixes in the drafting session's claim. The block records verdict: APPROVE (five minor points, applied) (block.yaml rereview_round_3), but "applied" is the drafting session's claim, not the reviewer's verdict. Round 2's record uses the same pattern. Fix (optional): verdict: APPROVE (five minor points) and fixes: applied in 95291e75.
• One long line. proposal.md:574 ("the house floor formula" line) now runs past the wrap width.

Lock and certification check.

• gates.yaml has 0 EPUF matches.
• The block has status: unlocked_report_only (:18), locked: false (:19), gated_cells: {} (:109) and lock_ceremony.exists: false (:216).
• The header says it "gates nothing" and "edits no gates.yaml byte".
• The diff adds no wording about passing, locking or certifying. The phrase "met the rules for gating" sits beside "report the comparison without a pass or fail".

What I could not check:

• Tests. I ran no tests. So I couldn't confirm that the two new tests pass, that the tier count of +2 (3,359 → 3,361) is right, or that the block equals a fresh render.
• Git. I ran no git commands. I took the delta from .diag/round3-fix-delta.diff, and the claim that the derivation-core diff since ce8d5000 is empty from the drafting session.
• Hashes. I computed none. The supplement and artifact hash literals are unchanged in the diff.
• PR body. I didn't read it.
• PSID data. I read no PSID microdata.

[MaxGhenis]

Merge record:

• Max's go: 2026-10-03 in chat ("yes 1/2" on d872).
• Merge: merge commit 492d5a6, pinned with --match-head-commit 358e15b5.
• CI: gh pr checks 509 exits 0 on 358e15b. MERGEABLE; marked ready before merge; no CHANGES_REQUESTED.
• Independent review (Opus 5.5 lanes):
  • rounds 1–4 are committed under reviews/; round 4 (20261002-173551-epuf-gate-r4) was APPROVE on 95291e7;
  • round 5 (20261003-220943-pr509-review-r5) was APPROVE on 5693804, nothing blocking. That commit applies round 4's three optional points and records round 4.
• After round 5:
  • 65daf03 applies round 5's only nonblocking point. The block's history now lists every review round, each keeping the verdict apart from the fix commit, and the rounds test checks this. A deliberate break of the history makes the test fail.
  • 358e15b merges master (Pin the depletion-cut outputs to the generator's own code, not all of src/ #514). The tier-count conflict was resolved from the three-way counts: unit 5,828 + 120 (Register gate_epuf against SSA's Earnings Public-Use File (unlocked, report-only) #509) + 12 (Pin the depletion-cut outputs to the generator's own code, not all of src/ #514) = 5,960; artifact 3,361; total 11,402. CI's tier-policy test passed on it.
• Report-only: nothing locks. gates.yaml has no EPUF entry, and the block keeps locked: false and gated_cells: {}.