Gate-2 chronic-cell forensics: pathway, parity, and draw-noise decomposition

[MaxGhenis]

Gate-2 chronic-cell forensics (reported, not gated)

The registered diagnostic of issue #42, comment 4913512779 — evidence, not another spec draw. It reuses candidate 8's fit/simulate machinery (scripts/run_gate2_candidate8.py, #93; chain-imports c1–c7) on the TRAIN side: for each gate seed it fits on side B and simulates the same side B, scoring against side B's own empirical rate. The outer holdout (side A) is never simulated — the only side-A numbers used are the already-published per-seed scores read from the committed runs/gate2_hazard_v8.json. gates.yaml and every committed gate artifact are untouched. reported_not_gated: true.

Findings

1 — Marriage-count pathway (mean_lifetime_marriages|male): the deficit is a residual the hazard model cannot represent, not a remarriage-level miss. Train-side, the simulation exceeds the reference in-exposure marriage count (male sim 1.365 vs ref 1.315; in-exposure deficit −0.050). The whole +0.065/person deficit is carried by the reference residual = 0.116/person — marriages counted in n_marriages with no datable/scoreable in-exposure event — which reconciles exactly to undatable_na_start 0.062 + n_marriages_minus_datable_episodes 0.040 + remarriage_prior_dissolution_year_undatable 0.013 (+ ~0 separation-origin). The largest datable per-cell deficit is a specific after_divorced | 2nd | age 25–34 cell (+0.014/person) that the origin-split remarriage cells pool away. This sharpens the c8 grading's "compositional, not remarriage-level": lifting remarriage hazards cannot close it — the model needs to represent the residual pathway.

2 — completed_fertility.c1970s: a broad low-parity progression miss, not a "share reaching 2+" miss, and not censoring. The simulation under-produces completed parity by 0.122 child/woman (ref 2.121 vs sim 1.999). Decomposing mean = Σ_k P(parity ≥ k), the per-margin deficits decline monotonically — 0→1 +0.036, 1→2 +0.035, 2→3+ +0.020, 3+→4+ +0.014 — so the two low-parity steps carry 58% of the gap; the model runs low at every birth, most at the first two. Not censoring: only 0.24% of reference cohort births fall outside the sim's [15,49] window (convention effectively shared). Not timing: mean mother age ref 27.90 vs sim 27.76.

3 — RNG stability (20 draws/seed, sim-RNG seeds 5200+k; the c10-diagnostics analogue). Every failing cell carries a small, reproducible train-side under-production tilt (same sign on all 5 seeds), so none is pure noise around a matched level — they split by how close the tilt sits to the tolerance:

cell	train tilt (signed)	tol	P(fresh draw clips)	verdict
mean_lifetime_marriages|male	−0.046	0.047	0.49	BOUNDARY
share_widowed.75+|female	−0.157	0.185	0.34	BOUNDARY
share_divorced.45-54|female	−0.103	0.212	0.03	NOISE-DOMINATED
widowhood.75+|female	−0.102	0.373	0.00	NOISE-DOMINATED
mean_lifetime_marriages|female	−0.036	0.051	0.10	NOISE-DOMINATED
completed_fertility.c1970s	−0.060	0.171	0.00	NOISE-DOMINATED

The outer FAIL 0/5 is largely draw/split luck around a model that runs slightly low across the board. The registration's knife-edge share_divorced.45-54|female (outer 0.212 clip decomposes to 49% level / 51% split+draw) and the seed-specific widowhood.75+ incidence (26% level / 74% noise) are noise-dominated — real but sub-tolerance tilts. Only mean_lifetime_marriages|male (on its boundary, clipP 0.49) and share_widowed.75+|female (85% of tol) sit close enough to be honest level targets.

Candidate-9 implications

Move the two boundary cells — the male marriage-count residual pathway (Q1) and the 75+ widowed stock — and the c1970s low-parity progression (Q2). Do not chase the noise-dominated knife-edges (share_divorced.45-54|female, widowhood.75+ incidence): their clips are split/draw excursions, not level misses. Register c9 only after citing this evidence, under the one-shot rule.

Delivered

• scripts/gate2_forensics.py — the runner (reuses c8 fit/simulate train-side; ~155 s per-seed compute over 5 seeds × 20 draws, plus load).
• runs/gate2_forensics_v1.json — the artifact (reported_not_gated, registration pointer, per-question blocks with every headline recomputable from stored per-seed values, per-seed + total compute time).
• tests/test_gate2_forensics.py — consistency tier (headline / residual-reconciliation / outer-score / tolerance / verdict recompute) + two skipif-PSID live pins (seed-0 train reference and first draw reproduced bit-for-bit).

Additive only; nothing existing modified.

🤖 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 8, 2026 10:59am