W1 candidate 1: first full transport deployment (registration 4950931131)

[MaxGhenis]

One-shot scored run of W1 candidate 1 — the first full transport deployment against the locked gate_w1 contract. Registered at issue #42 comment 4950931131. Publishes regardless of the verdict; no holdout tuning. The certified PSID-estimated generators (gate-1 earnings, gate-2a/2b marital CANDIDATE_16, gate-2b household CANDIDATE_9, gate_m4 disability, #115/#117 reform ledgers) are deployed on the certified populace CPS frame (us-4.18.8, sha c2065b64…, sha-verified) — a frame none was fitted to.

Verdict: FAIL (0 / 5 seeds)

gate_pass = false — family A 0/5 seeds, family B 0/10 cells, family C neither fingerprint reverses. This lands inside the registration's forecast P(gate ≥ 4/5) = 0.05–0.20. The value is the per-family decomposition and the fingerprint diagnostic.

Per-seed verdict table (family A, 53 gated cells)

seed	cells pass	cells fail	seed pass
0	28/53	25	❌
1	28/53	25	❌
2	32/53	21	❌
3	26/53	27	❌
4	24/53	29	❌

Conformance: identity_candidate = false, max across-draw sd 0.475 — the candidate regenerates every scored column per the regenerated_surface rule (not the prohibited identity map, which would score 0 with zero dispersion).

The fingerprint reversal results (family C — the highest-value read)

Both fingerprints were re-run on the transported frame changing only the frame (the committed #115/#117 ledgers run verbatim; the earnings-source readers are swapped to gate-1-transported careers). Neither reverses to the anchor ordering.

fingerprint	PSID "before"	required "after"	deployed	reversed?
C1 PPI↔NRA	PI > NRA > PPI > COLA	PI > PPI > NRA > COLA	PI > NRA > PPI > COLA	NO (τ vs required 0.667, unchanged from PSID)
C2 elim↔+2pp	+2pp > elim > +1pp > cap	elim > +2pp > +1pp > cap	elim > cap-$150k > +2pp > +1pp	NO (τ vs required 0.333)

• C1 does not move at all — PPI stays below NRA→70; the transported AIME above the second PIA bend does not lift progressive price indexing past the NRA cut.
• C2 shows the target swap but not a clean reversal — the required elimination > +2pp swap is realised (required_swap_realised = true): the heavier transported earnings tail lifts taxable-max elimination above +2pp, exactly the Roadmap: from incidence replication to a DYNASIM-class projection on one unified person-period panel #113 mechanism. But the same tail also lifts cap-$150k above both payroll rates, adding a second adjacency change, so the ordering is not the committed [elim, +2pp, +1pp, cap].
• Robustness: the two orderings were identical across three independent transport constructions (level-anchored, rank-flat, and rank + transitory mobility). The transported frame carries 37.7% of taxable payroll above the wage base — an upper read (positive prime-age earners, no zero/low-earning years), well above the real ~17% and the PSID frame's 12.7%. A heavier tail favours reversal, so the non-reversal is conservative: a realistic tail would be no more likely to reverse.

Per the registration, "their failing to reverse would be the single most important diagnostic in the program." The diagnostic here: deploying the PSID-fit generators on the CPS frame does move the taxable-max ordering in the predicted direction (C2 swap realised) but does not cleanly reproduce either committed anchor ordering.

Per-family decomposition

Family A (CPS-observable joints, floor-priced) — modal failure by view (cell-fails summed over 5 seeds):

view	cell-fails	reading
marital_share	45	younger cohorts (25–34, 35–44) under-produce married: the synthetic-panel marital deploy starts everyone never-married and the hazards do not equilibrate over the exposure window (older 45–64 cohorts largely PASS)
earnings_participation	25	gate-1 has no sex covariate and fits ages 25–59; 18–24 / 62–69 extrapolate to the nearest fitted bin, overstating retirement-age participation
hh_size_share	25	the household generator composes ADULTS; committed rate_a is all-person (children live in larger households) + coresident-parent/multigen are under-generated on a cross-section
coresident_spouse	16	tracks the marital under-production at younger ages
earnings_profile	16	the 18–24 age-boundary extrapolation

19 earnings cells pass (including the p90/p50 and p50/p10 dispersion ratios), so the earnings regeneration is not degenerate. (The registration predicted earnings-profile joints as the modal class; the actual modal class is marital_share, reported as-is.)

Family B (SSA anchors, 10 gated) — 0/10. DI status is M4-simulated (never the frame's DI column or any SS proxy; conditioning columns {age, is_female} asserted populated). The M4 work-disability prevalence gradient is far flatter than the SSA disabled-worker beneficiary stock's steep concentration at older ages: e.g. di_prevalence.60-fra deployed 18.6pp vs anchor 45.4pp; conversion share ~6pp vs ~14pp. The gradient transports; the level and steepness do not — the second failure class the registration named.

Family C — both fingerprints fail to reverse (above).

Spec resolutions (documented; the locked text does not pin them)

The certified generators are PSID-panel simulators; deploying them on a CPS cross-section required synthetic-panel adapters. Eight resolutions are recorded in the artifact's spec_resolutions and the module docstring, including: generators fit once on the full PSID sample and frozen (nothing refits per W1 seed); earnings regenerated from the certified gate-1 cell marginal at (age-bin, terminal period 2022), not anchored on the frame's own earnings (which would be the identity); marital/household via synthetic panels with empty initial roster states; family-A surface scored on the adult universe (only hh_size_share's committed all-person rate_a differs in scope, disclosed); DI via Bernoulli(m4_prevalence[band,sex]); family-C careers transported with the committed permanent-variance share (0.467). One label note: gates.yaml calls the household generator "gate-2c" at the hh_size_share row; household composition is the gate-2b-locked CANDIDATE_9 everywhere else in the file, used here.

Reproduction

# 1. data-bound frame export (policyengine.py .venv, sha-verifies c2065b64…)
PYTHONPATH=src <pe.py-venv>/python scripts/export_frame_persons.py
# 2. fit + score (.venv-gate); family A chunked by seed per the memory guard
POPULACE_DYNAMICS_FRAME_PICKLE=scratch/frame_persons.pkl \
  <gate-venv>/python scripts/run_gate_w1_candidate1.py fit
#   then: seed 0..4 (parallel) + family-b + family-c, then assemble

Runtime: generator fit ~2 min (gate-1 marginals + ft 19s + hc 82s + m4); family A ~5 min/seed (5 parallel processes); family B ~2s; family C ~60s. Frame 166,302 persons / 340.0M weighted, sha-verified against the pin.

Tests

+11 tests (9 unit pure-logic, 2 artifact committed-json bindings — including the holdout-split binding that recomputes every gate seed's committed household sha256 with no h5). tier_counts.json refreshed in a CI-equivalent env (.venv, openpyxl present) per the #158 rule: unit 199→208, artifact 783→785. Always-runnable tiers green: 953 passed, 40 skipped (unit or artifact); test_tier_policy green over the full 2451-test collection. black + ruff clean.

Do not merge on my account; verification-before-record.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
social-security-model	Ready	Preview, Comment	Jul 12, 2026 12:38pm