Gate-2 candidate 9 (run 1): FAIL 0/5 — observed undatable-marriage count init + unsmoothed low-parity fertility

[MaxGhenis]

Verdict: FAIL — 0/5 seeds pass

Pre-registered one-shot run of gate-2 candidate 9, frozen spec issue #42 comment 4914111252. Published regardless of verdict; not merged.

Candidate 8's spec (scripts/run_gate2_candidate8.py, #93) with exactly two deltas registered off the gate-2 forensics (#94):

1. Observed undatable-marriage lifetime-count initial state (delta 1, RNG-neutral). Each holdout person's simulated lifetime-marriage count initializes at their observed residual R = n_marriages − datable in-exposure first_marriage/remarriage events — the forensics Q1 residual (undatable start/dissolution, out-of-window/underage, separation-origin, MH-count-excess) at per-person resolution — then accumulates the simulated datable transitions. Hazards untouched; no draw perturbed.
2. Unsmoothed low-parity peak fertility (delta 2). Parities 0→1 and 1→2 use the exact empirical single-year rate (no kernel) at ages 22–38 where cell exposure ≥ 200 weighted person-years; kernel elsewhere; higher parities unchanged.

Seed conjunction

seed	gated pass	seed_pass
0	44/46	False
1	45/46	False
2	43/46	False
3	45/46	False
4	45/46	False

Gate rule: all 46 gated cells per seed AND ≥ 4 of 5 seeds. Result: 0/5 seeds → gate_2_pass = False.

Per-block pass counts (per seed)

block	cells	s0	s1	s2	s3	s4
divorce	4	4	4	4	4	4
fertility	11	10	11	10	11	11
first_marriage	6	6	6	6	6	6
nuptiality_cohort	5	5	5	5	5	5
occupancy	6	6	6	4	6	5
remarriage	4	4	4	4	4	4
stock_occupancy	6	5	5	6	6	6
widowhood	4	4	4	4	3	4

Every failing gated cell (score / tolerance / seed)

seed	cell	score	tol	c8 score	c8 pass
0	asfr.20-24	0.1086	0.1010	0.0956	True
0	`share_divorced.45-54	female`	0.2123	0.2120	0.2123
1	`share_widowed.75+	female`	0.2059	0.1850	0.2059
2	completed_fertility.c1970s	0.1983	0.1710	0.1863	False
2	`mean_lifetime_marriages	female`	0.0609	0.0510	0.0015
2	`mean_lifetime_marriages	male`	0.0510	0.0470	0.0244
3	`widowhood.75+	female`	0.3915	0.3730	0.3915
4	`mean_lifetime_marriages	female`	0.0565	0.0510	0.0076

Delta-1 reconciliation (verified before the one-shot)

• Per-person identity R + datable == n_marriages: max abs residual 0.0 (remainder 0.0).
• R aggregated over the ever-married denominator == forensics reference residual == five-bucket sum: max abs remainder 2.91e-16 across all 5 train seeds × 2 sexes.
• Residual non-negative (min 0.0), integer-valued; 403 persons with undefined n_marriages carry R = 0.
• Matches the committed forensics per-seed buckets bit-for-bit (runs/gate2_forensics_v1.json).

Byte-identity to candidate 8

Delta 1 adds no draw and delta 2 moves only the fertility threshold (parity-independent draw sizes), so the marriage draw stream is identical to candidate 8. The 165 non-fertility, non-count gated cells (33 cells × 5 seeds) match candidate 8 to 0.0 deviation; only the fertility cells (delta 2) and the two lifetime-marriage counts (delta 1) move.

Movement vs candidate 8 (n_seeds_pass)

cell	c8	c9	note
`mean_lifetime_marriages	male`	3/5	4/5
`mean_lifetime_marriages	female`	4/5	3/5
completed_fertility.c1970s	4/5	4/5	delta 2 did not clear it (seed 2, worse score)
asfr.15-19	5/5	5/5	unaffected
asfr.20-24	5/5	4/5	delta 2 broke it on seed 0
`share_divorced.45-54	female`	4/5	4/5
`share_widowed.75+	female`	4/5	4/5
`widowhood.75+	female`	4/5	4/5

Reading. Delta 1 raises mean_lifetime_marriages|male (3/5→4/5, curing seeds 3–4) but overshoots where the hazards already over-produce in-exposure marriages (the forensics warning), dragging mean_lifetime_marriages|female down (4/5→3/5, failing seeds 2 and 4). Delta 2 does not clear completed_fertility.c1970s and breaks asfr.20-24 on seed 0. The three byte-identical elderly-widow / divorce-stock cells the registered modal named (share_widowed.75+|female, widowhood.75+|female, share_divorced.45-54|female) hold exactly where candidate 8 failed them — no delta can touch them — and are the structural blockers on seeds 0/1/3.

Discipline

• Bit-exact precheck: reproduced the committed floor (ref dev 0.0, rate_a dev 0.0, holdout sha256 all match).
• One outer run; per-seed cache outside runs/; published regardless.
• Artifact runs/gate2_hazard_v9.json (registration pointer, both deltas, reconciliation record, forecast pointer).
• Tests tests/test_gate2_candidate9.py (consistency + delta-1 reconciliation + delta-2 footprint + byte-identity to c8 + live seed-0 pin + structural delta checks).
• black -l 79 + ruff clean. c1–c8 + forensics tests stay green; full suite 972 passed, 41 skipped.

Pre-registered forecast was P(pass) ≈ 0.40–0.50 with share_widowed.75+|female as the modal blocker; the actual decider was broader — the byte-identical-to-c8 marital cells plus delta 1's overshoot held the gate below 4 passing seeds. 🤖 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 12:22pm