M6 factory: (0,24) projection-coverage pad + pe-us param-dir binding (3e patch: F1/F2/N1 + bridge test)

[MaxGhenis]

The zero-discovery patch pinned by ratified amendment 3e (§2.8.10.5, merged #196) and verified against the merged engine by the 3e referee (#196 comment). Every design decision here was made and refereed in the amendment; this PR only transcribes it.

F1 — the inert (0,24) projection-coverage pad

_pad_below_25_projection_coverage(external_rates, exposure, *, boundary_year) appends one (0,24) band per sex to both mortality fit inputs, after the seven-band construction (build_inputs step 4):

• external pad: NCHS 2010 (l₀−l₂₅)/(T₀−T₂₅) central rate (≈0.000766 male / ≈0.000451 female) — provenance honesty only; aligned_rate = central_rate × (psid_rate/central_rate) = psid_rate cancels it.
• exposure pad: dated event_year = required_interview_year = T* (survives the ≤T* flow truncation), start_weight = exposure = 1.0, death = 0.0 → the fitted <25 hazard is −expm1(0) = 0 exactly: no modeled mortality below the age-25 PSID exposure floor.

MORTALITY_BANDS, build_mortality_floors.BANDS, and the committed floors artifacts are untouched; the MORTALITY_BANDS == mf.BANDS guard still binds.

F2 — the parameter-dir binding

assert_pe_us_param_dir() (step 1, right after assert_pe_us_version()) requires (_resolve_pe_us(None)/_SSA).resolve() — the directory load_ssa_parameters will read — to equal the metadata-versioned distribution's own parameters/gov/ssa. Closes the referee-demonstrated silent-wrong-load seam (1.752.2 metadata + a 1.690.7 checkout behind the env var).

N1

The extractor's era-mapping docstring no longer claims a build() verification that does not exist; era records are provenance-only, read by nothing.

The missing bridge test (referee-flagged)

Committed tests now route factory-shaped rates through prepare_mortality_refit_inputs → fit_mortality_model — the run's exact first phase:

• unpadded seven-band inputs raise mortality bands must start at age zero (the failure the pad exists to prevent);
• padded inputs fit a valid 8-band 0→120 model; <25 hazard exactly 0 both sexes; the fourteen 25+ cells equal the hand-computed PSID arithmetic −expm1(−deaths/exposure) and are invariant to pad value (0.5) and weight (17.3);
• pad rows verified against the NCHS lx/Tx formula and the ≤T* truncation;
• param-dir gate tests run against an injected fake distribution — no real policyengine-us install needed, so the module stays in the always-run unit tier (the env var is referenced through the loader's own _PE_US_ENV constant; a literal would have flipped all 31 module tests to the oracle tier and silently demoted 21 existing always-run factory tests).

tier_counts.json: unit 410 → 420 (CI-equivalent full-collection recount; tier policy green).

One disclosed addition beyond the pinned pointers

_KNOWN_STALE_REGISTRATIONS gains registration 2 (4967241464), whose designed pre-scoring stop is publicly graded (comment 4967433717) — the runner guard must not accept a superseded registration for a PSID-touching run. Guard hygiene from the public record; no design choice, no data selection. (tier_counts.json artifact 1010 → 1011 for the added parametrize case.)

Not in scope

No gates.yaml change, no floor, no threshold, no scored surface. After merge: the THIRD gate_m6 registration on #42, then the one-shot temporal-holdout run.

🤖 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 14, 2026 3:34pm