M6 design amendment 3: the scored-run harness (§2.8, unblocks the gate_m6 run)

[MaxGhenis]

What

Adds §2.8 — the scored-run harness to docs/design/m6_projection_engine.md (design amendment 3). Docs-only: it edits no gates.yaml cell, moves no threshold, builds no floor, and writes no test.

Why

gate_m6 candidate-1 was registered as a one-shot temporal-holdout run (#42 comment 4962640241) and then stopped before any scored phase — the designed pre-scoring stop, comment 4962773701, graded a registration error. The engine (PR #173, merged 95a3e05) is build-only: the per-slice native-panel builders exist only as test stubs, the drift-scoring composition is unbuilt, and pre-flight 1 had only run on synthetic ensembles. The stop names this design amendment as the unblock path (referee round → build → engine-referee-style review → merge → fresh registration).

§2.8 pins every modeling decision so the harness build lane implements with zero design choices (the §2.7.6 bar), candidate-blind (builders and scorer exercised on synthetic frames only until the registered run; no holdout truth read in the build).

The pins (§2.8.1–§2.8.9)

• 2.8.1 the adjudicated core — the year-0 closed panel carries realized histories: the builders read realized ≤2015-interview marriage/household histories for the year-0 population; projected years are constructed from simulated state. Argued from decision 5 + §4.4 + the built disability_step reproduction template (assembly.py:257-304) + support.py's GATED_REALIZED/FORWARD modes. Forward/production path out of scope (decision-5 successor gate); pointer restated.
• 2.8.2 the projected-slice→native-panel builders, per field — against the certified MaritalPanel (episodes→changepoints→backward-asof) and HouseholdCompositionPanel (observed-states→evolve) schemas as the contract: episode extension, changepoint appends, duration accounting, censor handling — each stated in the certified schema, R1 plumbing, whole-window core run once per draw (the disability_step cache pattern).
• 2.8.3 the year-0 realized-2015-interview slice — mirrors build_anchor_frame / presence_by_wave and the floor's PSID reads exactly (SEED_WAVE=2015); F6 start-wave weight and household_id fixed; presence-conditioned symmetrically via prepare_evaluation_support.
• 2.8.4 the M6 drift-scoring layer — reuses the floor cell functions verbatim (named: build_anchor_frame, presence_by_wave, *_slices, *_cells, _rate, _wquantile, _score, split_panel_by_person, derive_tolerance), with a byte-identity self-test if reimplemented; per gate seed → side-A half-split, K=20 rbar vs realized rate_a, |ln(rbar/rate_a)| against the frozen v3 floor (e931c886…), 4-of-5 seed conjunction read from the locked block; undefined-draw + regenerated-surface guards.
• 2.8.5 pre-flight 1 — the candidate-9 re-certification margin on the ≤2014-refit real panel over the gate seeds (check_candidate9_recertification, ≥3σ), holdout-blind; failure = designed abort (one-shot not consumed).
• 2.8.6 pre-flight 2 — certified sign-path verification (_gate_sign_draw draw_sign branch vs the test seam), synthetic probe, path recorded.
• 2.8.7 the runner phases — refit | pre-flight×2 | project+score per seed | report-only | assemble; one-shot write_new + .env.json sidecar, stamping the registration-id + EARNINGS_SPEC_REGISTRATION.
• 2.8.8 the fence, extended — no gates.yaml read beyond the block's protocol/cells; no holdout-informed choice; no realized post-boundary macro read on the scored path (the §2.7.6.3 leakage prohibition); forward-mode inputs stay rejected.
• 2.8.9 — residual open decisions: none. Two mechanical build-lane alignments recorded (once-per-draw cached core; extract the floor cell functions to one shared module).

Also updates the revision header (→ revision 4 / amendment 3) and the §10 design-parameters block (adds scored_run_harness).

Verification

• docs-only — no engine, floor, or test changed.
• All 39 design-doc-bound tests pass unchanged: pytest tests/test_gate_m6_floors.py tests/test_gate_m6_derivations.py → 39 passed.
• The §10 m6-design-parameters JSON block parses.

Draft pending the amendment referee round.

🤖 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 13, 2026 11:13pm

[MaxGhenis]

Amendment-3 referee round (§2.8, the scored-run harness): MAJOR REVISION

One finding of the §2.7.6 class — a pin that contradicts the certified code on the gated surface — plus five zero-choices precision gaps. Everything else in §2.8 verified code-true, line-exact, and certification-consistent; the revision is scoped, not structural. Verified against 22e8af0 (worktree build, .[dev] venv, PSID staged): every cited line was opened, the floor/gate/engine machinery was read end-to-end, and the harness was mentally implemented phase by phase against loop.py/assembly.py/support.py/composition.py/steps.py/forward_earnings.py/refit.py, the v1/v3 floor scripts, and the locked gates.gate_m6 block.

---

MAJOR

1. The gated-earnings scored support is contradicted by certified step-1 mortality — the one gated family §2.8 leaves unresolved.
§2.8.3 pins presence sets "applied to projection, truth, and floor symmetrically by support.prepare_evaluation_support," which raises on any support mismatch (support.py:311-315, "symmetric presence-conditioning requires identical projection and truth person-period support"). §2.8.4 pins the projected side as a reduction of "each K-draw's projected panel" — the engine's frame slices (loop.py:136-143). But the certified loop's step 1 removes simulated deaths from every subsequent frame (apply_mortality, steps.py:108-130, "return only the period's survivors"), so the projected earnings person-periods at {2016, 2018} are a strict subset of the realized truth support (family_earnings_panel rows, earnings_frame, build_m6_holdout_floors.py:354-378) on essentially every draw — dozens of simulated-dead-but-realized-present persons per draw at 25–64 mortality over 1–4 projected years, per side-A half of the 29,792-person closed panel.

The amendment resolves exactly this interaction for the other three gated families — marital via the realized presence/death censor on a frame-independent panel (§2.8.2; the builder reads the §2.8.3 realized sources, not the live frame), disability via realized-support reproduction (assembly.py:257-304), household via realized person_waves support — and is silent for earnings, the only gated family computed from the live frame. As pinned, the build lane lands in one of two unadjudicated states:

• (i) if the earnings frames route through the §2.8.3 guard, the run aborts on every draw (and note presence_by_wave's odd-wave keys cannot even key the even reference years — the earnings-side conditioning machinery is unspecified, see finding 6); or
• (ii) if they do not, the run silently scores survivor-conditioned projected earnings against all-present realized truth — an unpinned support decision on 6 of the 11 gated cells, with a systematic direction (mortality–earnings correlation biases earn_p10.prime, earn_zero_rate.older et al.), and it functionally admits the mortality module — not_certified mortality drift, first margin — into the gated surface through survivorship.

Either branch falsifies §2.8.9's "residual open decisions: none." This is the analogous class to the §2.7.6 fit-range/gated-surface MAJOR: a scored-surface support definition the certified code contradicts. The ratified record points at the resolution — §4.7 calls the construction "the gate_m4 construction with the time axis substituted," and the gate_m4/reproduction precedent scores on realized support; §4.4's cancellation logic ("the same wave-presence selection sits on both the projected and the realized rate") requires it — i.e., the scored earnings person-periods should be the realized support, with the forward chain evaluated for realized-present persons regardless of simulated death. That is cleanly implementable without perturbing survivor bits (apply_earnings draws per person-keyed streams, steps.py:224-233, so evaluating dead persons' chains consumes no survivor entropy), and it keeps mortality fully report-only. But choosing it — versus disclosed survivor-conditioning — is an adjudication + pin, not build-lane mechanics. Amend §2.8.4 (and §2.8.3's presence bullet) to pin the gated-earnings support explicitly.

Required pins (zero-choices gaps; one sentence each)

2. The once-per-draw cached cores' stream address is unpinned. The certified cores are per-trajectory, one stream set per draw (verified — see the record below), but §2.8 never says which registry period the single cached run consumes: the first-invocation period-1 streams the loop passes (loop.py:234-237; assembly.py:213-215), or the period-0 deployment/compatibility address that reproduces the certified topology (rng.py:41-44; composition_rngs_from_registry(…, 0), composition.py:86-93 — the address the §2.8.5 internal reference itself uses, composition.py:643-653, n_periods=0). Both are candidate-blind and statistically valid; they differ bit-wise, and §4.9 pass verification is bit-exact. §2.8.2's household sentence — "consuming … the per-period HOUSEHOLD_COMPOSITION streams" — is additionally self-contradictory for a single whole-window call, which takes exactly one CompositionRngs (simulate_candidate9_injected, composition.py:532-539). Pin the address (and the marital main_rng/gap_rng and step-4 fertility stream with it).

3. The marital builder's person universe is under-pinned. The §2.8.2 table sources attrs "from the §2.8.3 year-0 slice" with holdout_ids = the full closed panel — but anchor persons absent from the marriage-history file (children; MH-uncovered adults) have no truth-side marital exposure at all (person_attributes/_valid_persons, transitions.py:239-277), while a builder that seeds them projects them as never_married at-risk person-years (the pd.isna entry path, marital.py:100-101) — asymmetric first_marriage denominators for the 16–17-at-2015 cohort entering band 18+ in-window. The entry-row cell's "identical machinery the truth side runs" implies the right universe; make it explicit: marital attrs = build_marital_panel(mh, deaths, anchor_weight).attrs ∩ anchor (which also keeps start_exposure ≥ birth_year + START_AGE and resolves the missing-entry-row case by construction).

4. "Realized ≤2015-interview histories" is wrong for the 2017/2019 openers. §2.8.1's headline and the §10 JSON core_adjudication pin the seed boundary at the 2015 interview; the 2017/2019 presence-conditioned openers (§2.8.3, correctly) seed at their anchor wave — their realized ≤2017/≤2019-interview state, which reads realized 2015–2018-dated events into the seed. That is the ratified conditioning (block f6_weight.start_wave; §4.4), not a leak — but the boundary statement should read "≤ the person's anchor interview (2015 for the bulk)" in §2.8.1 and the JSON, so the leakage fence is stated at the true per-person boundary.

5. The attrs.n_marriages pin is internally inconsistent (inert, but a pin should be exact). "Realized marriage count as of the seed wave … via person_attributes" — person_attributes yields MH18, the lifetime count as of last report (transitions.py:239-243; marriage.py:216), which includes post-boundary marriages; those differ. The field is provably unread by the core (order re-seeds to 1, marital.py:107,119) and overwritten in the output (_assemble_panel recomputes it from simulated episodes + residual, simulator.py:76-87, marital.py:262-272) — say which value the builder writes and that it is carried-inert, rather than pinning two different values in one cell.

6. Name the earnings-side support basis. §2.8.3's presence bullet covers flows only (presence_by_wave, START_OF_INTERVAL). The earnings truth support is the realized panel's own person-period row existence at {2014, 2016, 2018} (earnings_frame; "present, positive-weight, valid-earnings"), an EXACT_WAVE-class basis on reference years. One sentence naming it — folded into finding 1's resolution — closes the gap that lets a builder mis-apply the flow presence sets to even reference years.

Notes (disclosure-grade, no gate impact)

• Fertility remains whole-window-recompute-per-period (apply_fertility → simulate_fertility on the cached marital result each period with the period stream, steps.py:401-433): frame-materialized births are stitched across years. Inert on every gated cell (children are outside holdout_ids and the cell bands; births feed report-only entrants + the household roster) — worth one disclosure line or a third mechanical alignment.
• The 2c-modifier view cannot move the gated marital cells: apply_first_marriage_modifier is (band, sex)-marginal-preserving by construction (§2.6 point 3b), and the scored reduction reads the raw panel events.

Verified sound (the record this round confirms)

• §2.8.2 per-field pins are code-true, line-exact: seed mapping current_start = start_year − duration, dissolution_year = start_year − years_since, order re-seed (marital.py:104-119); loop over [start_exposure_year, censor_year] (marital.py:150-153); duration/years-since verbatim (marital.py:188,217); intact-at-censor end=NA (marital.py:243-249); kind_order dissolve-before-marry + allow_exact_matches=False (transitions.py:307-352,377-423); the simulated panel re-assembled through the certified _changepoints/_assign_state/_events_frame (simulator.py:67-118); no new changepoint kind. Household: the pinned seed states are exactly the consumed inputs (obs_parent/obs_multigen composition.py:240-241, obs_cohab :302, obs_skipgen :475-479), the evolve kernels read only wave-0 seeds (common.py:87-130), _add_transitions rebuild matches the drop/re-add (composition.py:584-593).
• Censor symmetry (the P6-class check): SYMMETRIC. The pinned projected censor min(2022, person_attributes censor) equals the truth panel's min(mrr, death, 2023) everywhere inside the 2015–2021 scored/shock window; universe (anchor ∩ MH-valid, modulo finding 3), presence sets (keep_present's opening-wave rule = presence_by_wave), and F6 weights are the same objects on both sides; support.py:311-315 backstops the identity at runtime. Attrited-not-dead, absent-mid-window, and 2017/2019-opener cases all trace symmetric.
• The adjudicated core's precedent is real and extends without new leakage: disability_step builds reproduction support from realized inputs.disability_panel filtered to [start_year+1, end_year] — not the projected frame — and caches one simulate_reproduction run (assembly.py:257-304); GATED_REALIZED requires / FORWARD structurally rejects realized inputs with the named structural_delta (support.py:236-264). Nothing post-boundary-dated enters any seed (backward-asof excludes boundary-year changepoints); every post-boundary read the builders make (censor, presence, anchor, F6) is the ratified §4.4/§4.7/block conditioning set — no addition.
• Once-per-draw is certification-consistent; per-period recompute is not (the flagged §2.7.6-class hunt — it resolves in the amendment's favor): the certified candidate-16 is one whole-window trajectory per draw on one main + one gap generator ("preserve run_gate2_candidate16.py:826-1079 exactly", simulator.py:121-133; §3.1's K=20 = twenty whole-window runs on 5200+k), and certified candidate-9 is one whole-window evolve per draw at the frozen tag topology (composition.py:86-111,635-653). The built loop invokes the cores per period with per-period streams (loop.py:234-274) — with a real whole-window builder that stitches independent draws across years and breaks state-follows-events in the longitudinal marital cells, exactly as §2.8.2 argues. The cache is the certified shape; only its stream address needs the finding-2 pin.
• Drift scoring: every named reuse target exists as named (build_anchor_frame :153-171, presence_by_wave :174-181, marital_tables :252-299 with the :259-261 machinery line-exact, mortality_slices, disability_pairs, earnings_frame, the four cell functions, MARITAL_AT_RISK, EARN_COHORTS, band_of, _rate, _wquantile, _score :603-621, harness.moments, split_panel_by_person, derive_tolerance); _score's metric branches are exactly the pinned log-ratio/abs-gap split; the byte-identity self-test is the right and sufficient fidelity pin for the floor-bound quantities (rate_a, tolerances), with extraction making identity by construction; gate seeds {0–4}, the 4-of-5 conjunction, the 11 cells, split_units, metrics/k, and tolerances are all read from the locked block (verified against gates.yaml gate_m6 views/scoring), and K=20 on 5200+k restates the ratified §3.1/§4.7–4.8 pins — invented nowhere; the dispersion/conformance guards match the prior scored-run artifact schema.
• Pre-flights + fence: §2.8.5 is §2.6 requirement-4 verbatim — output channel moments over RECERTIFICATION_CHANNEL_SETS, ≥3σ with no tolerance floor, raise-on-fail naming the fuller re-ceremony (check_candidate9_recertification, composition.py:692-758), injected-vs-frozen-ft.simulate reference, holdout-blind; §2.8.6's two paths are exactly forward_earnings.py:815-819 (draw_sign) vs :820-826 (_target_models seam). The fence's new-read-path audit holds: the generator provably never reads the frame's realized nawi (explicit guard + FRAME_COLUMNS-only selection, forward_earnings.py:1059-1061), fit-time NAWI is ≤2014-validated (_precutoff_nawi :209-229; ols_log_nawi_2005_2014, refit.py:98), odd years carry forward (:1051-1053), and advance_age's realized-macro columns (steps.py:143-156) are scoped to the admissible non-scored/report-only paths — the sole realized-post-boundary influence reachable into a gated cell is finding 1's survivorship channel.
• Runner/one-shot plumbing: refit_m6_components/M6RefitInputs/M6RefitBundle/RefitProvenance/EARNINGS_SPEC_SHA256 (refit.py:100-199,1063), EARNINGS_SPEC_REGISTRATION (refit.py:70), write_new exclusive-create + .env.json sidecar (artifacts.py:45-79), stub citations exact (tests/test_m6_engine_assembly.py:75-76,269-270).
• Hygiene: docs-only diff (one file, +390/−5); 39/39 doc-bound tests pass in a fresh venv (test_gate_m6_floors.py + test_gate_m6_derivations.py); the §10 m6-design-parameters JSON parses (revision 4); CI green on fresh runs; MERGEABLE/CLEAN.

Disposition: adjudicate + pin finding 1, add the five one-line pins (2–6), and this amendment meets the §2.7.6 bar; the flows/disability scoring layer, the adjudicated core, the pre-flights, and the fence need no structural change. No gate cell, threshold, or floor artifact is touched by any of the above.

[MaxGhenis]

Revision 1 — addresses the amendment-3 referee round (comment 4963629234: MAJOR REVISION)

Pushed 45e757d (docs-only, +119/−37). One MAJOR (F1) adjudicated + pin, the five zero-choice pins (F2–F6), and the fertility disclosure. §10 JSON → revision 5. All 39 design-doc-bound tests pass unchanged.

F1 (MAJOR — gated-earnings scored support vs certified step-1 mortality) → §2.8.4 + §2.8.3. Certified apply_mortality (steps.py:108-130, "return only the period's survivors") drops simulated deaths from every later frame slice, so the projected {2016, 2018} earnings person-periods are a strict subset of the realized earnings_frame support — the §2.8.3 identity guard (support.py:311-315) aborts every draw, or 6/11 cells score survivor-conditioned (admitting not_certified mortality drift through survivorship). Adjudicated, by the same symmetric-conditioning principle §2.8.1 applies to histories, §4.4's cancellation logic, and §4.7's "the gate_m4 construction with the time axis substituted" (whose reproduction precedent scores on realized support): the gated earnings cells are a reproduction read on the realized earnings support — the forward chain is evaluated for every realized-present person-period regardless of simulated death. Bit-safe because apply_earnings draws on person-keyed EARNINGS streams (steps.py:222-233), so a realized-present-but-simulated-dead person's chain consumes only its own entropy and perturbs no survivor's bits; projected mortality stays report-only. Stated plainly what it does/doesn't condition on (realized reference-year presence = conditioning-not-leakage; mortality's earnings-composition survivorship channel is not scored — the reproduction test's designed scope, same as the flows).

The support-basis sentence, verbatim:

The gated earnings cells score on the realized support, not the live projected frame (adjudicated — the one gated family §2.8 must resolve against certified step-1 mortality).

F2 → §2.8.2. Pinned the cached-core stream address to the period-0 deployment address (rng.py:41-44) — the certified single-period topology, and the address the §2.8.5 internal reference already uses (composition.py:643-653, n_periods=0) — with the marital main_rng/gap_rng and the step-4 fertility stream; fixed the "per-period HOUSEHOLD_COMPOSITION streams" self-contradiction → the single composition_rngs_from_registry(registry, 0) CompositionRngs per draw.

F3 → §2.8.2 table. Marital attrs = build_marital_panel(mh, deaths, anchor_weight).attrs ∩ anchor — the identical truth-side universe; MH-absent anchor persons (children, MH-uncovered adults) excluded on both sides symmetrically.

F4 → §2.8.1 + §2.8.2 household seed + §10 JSON. "realized ≤2015-interview" → "realized ≤ the person's anchor interview (2015 for the bulk; 2017/2019 openers seed later)" — the leakage fence stated at the true per-person boundary (ratified conditioning, no fitter sees the seed).

F5 → §2.8.2 table. attrs.n_marriages = the lifetime MH18 person_attributes writes (the truth-side value → byte-identical seed), carried-inert — the core never reads it (order re-seeds to 1, marital.py:107,119) and _assemble_panel overwrites it (marital.py:262-272). One value, not two.

F6 → §2.8.3. Named the earnings-side support basis — realized family_earnings_panel person-period row existence at {2016, 2018} (EXACT_WAVE-class on reference years), distinct from the odd-wave flow presence_by_wave sets; folded into F1's resolution.

Disclosure note → §2.8.9. Added the fertility whole-window-recompute-per-period disclosure (frame-materialized births stitched across years; inert on every gated cell — children outside holdout_ids and every band, births feed report-only entrant/roster only) as a third mechanical alignment.

No gate cell, threshold, or floor artifact touched. Stays DRAFT pending the referee's re-read.

[MaxGhenis]

Amendment-3 re-check (§2.8 revision 1, 45e757d): VERIFIED — ready to merge

All six findings are resolved, not renamed — each traced against the certified code, not the revision narrative. The MAJOR (F1) is adjudicated with a bit-safety claim I verified holds at both the entropy and the value level; F2–F6 are exact pins; the fertility disclosure landed. Verified on the worktree build (.[dev] venv, PSID staged): docs-only (+119/−37, one file), 39/39 doc-bound tests pass, §10 JSON parses at revision 5, CI green on 45e757d, MERGEABLE/CLEAN.

F1 (MAJOR) — the gated-earnings support adjudication and its bit-safety claim: VERIFIED from code.

The pin — gated earnings = a reproduction read on the realized family_earnings_panel support, the forward chain evaluated for every realized-present person-period regardless of simulated death — is bit-safe, and I confirmed the claim at the level that matters (does a dead-but-present person's chain perturb any survivor's value, not merely its entropy):

1. Entropy is person-local. apply_earnings's registry branch draws each person on context.person_generator(EARNINGS, person_id) (steps.py:220-224) → person_ordinals[person_id] → child_generator(period, EARNINGS, ordinal) (loop.py:91-105, rng.py:158-169) — a pure function of (k, period, module, ordinal), independent of frame composition.
2. Survivor addresses are provably invariant. person_ordinals is built once from the initial slice's sorted ids (loop.py:202-207) and is append-only (loop.py:283-285); mortality-dropping never removes an ordinal. So every survivor's stream address is identical whether or not the dead-but-present persons are carried in the earnings-scoring support — the exact property the reproduction read needs.
3. The value is person-local too — the half a rename would have missed. In the registry branch generate is called on a single row (frame.iloc[[index]], steps.py:225-226), and the generator's donor pools are frozen at fit time from the ≤2014 estimation panel (_build_donor_pools baked into ForwardEarningsGenerator(pools=pools), forward_earnings.py:762-792) — not rebuilt from the live frame; generate derives its seed solely from the passed per-person rng and reads only that row's conditioning + the frozen pools + the per-row gen_earn_w2/w4 chain memory (forward_earnings.py:1046-1107, steps.py:237-247). No live cross-person donor pool exists to perturb.

So evaluating a realized-present-but-simulated-dead person's chain consumes only that person's own entropy and reads only that person's own row + frozen pools — it changes no survivor's value or bits. The scored earnings quantity is therefore invariant to whether the implementation carries dead persons in a parallel earnings-support frame or a non-dropping frame — no residual scoring freedom (zero-choices re-check on the F1 path: clean). The support-basis bullet (§2.8.3) is correct — realized family_earnings_panel reference-year row existence at {2016, 2018}, EXACT_WAVE-class, explicitly distinct from the odd-wave presence_by_wave flow sets (whose keys cannot address even reference years). The scope statement is coherent with the locked block: mortality's earnings-composition survivorship channel is not scored, projected mortality stays report-only (feeds diagnostics + AIME/claiming/roster, not the gated read), and mortality drift stays not_certified — consistent with gates.gate_m6 (ZERO mortality cells gated, not_certified mortality_drift first).

F2–F6 + disclosure (resolution, one line each)

• F2 — period-0 address pin: VERIFIED. Cached core pinned to generator(0, MARITAL_CORE) / child_generator(0, MARITAL_CORE, 1) / single composition_rngs_from_registry(registry, 0) / generator(0, FERTILITY) — period 0 reproduces the certified SeedSequence([5200+k, tag]) topology (rng.py:117-119) and is the exact address the §2.8.5 internal reference already uses (composition.py:643-653, n_periods=0), so pre-flight 1 compares like-with-like; the "per-period streams" self-contradiction is gone (simulate_candidate9_injected takes exactly one CompositionRngs, composition.py:532-539).
• F3 — attrs ∩ anchor: VERIFIED symmetric. attrs = build_marital_panel(mh, deaths, anchor_weight).attrs ∩ anchor is the identical universe the truth side scores (marital_tables inner-merges anchor onto the same build_marital_panel, build_m6_holdout_floors.py:259-299); MH-absent persons drop on both sides via _valid_persons (transitions.py:265-277), which also enforces start_exposure ≤ censor and gives every seeded person an entry row.
• F4 — anchor-interview boundary: VERIFIED. §2.8.1 body, §2.8.2 household seed, and §10 JSON core_adjudication now read "≤ the person's anchor interview (2015 for the bulk; 2017/2019 openers seed later)"; the two remaining §2.8 "2015-interview" strings are the §2.8.3 title (line 954) and the SEED_WAVE = 2015 fact (line 956), whose body (lines 960-966) explicitly handles the 2017/2019 openers — genuine bulk-case, not a suppressed opener.
• F5 — n_marriages: VERIFIED one value. Now pinned to the lifetime MH18 person_attributes writes (transitions.py:239-243, marriage.py:216) = the truth-side value (byte-identical seed), carried-inert — the core re-seeds order to 1 (marital.py:107,119) and _assemble_panel overwrites it from simulated episodes + residual (marital.py:262-272). The two-different-values inconsistency is gone.
• F6 — earnings support basis: VERIFIED named, folded into F1's resolution (see above).
• Disclosure — VERIFIED. Fertility per-period recompute (apply_fertility → simulate_fertility, steps.py:401-433, births stitched across years) added as the third mechanical alignment; inert on every gated cell (children are outside holdout_ids and every gated band; births feed only report-only entrant/roster paths) — correct.

Non-blocking nicety (no gate impact, not a condition of merge)

The amendment-3 revision header bullet (line 18) still glosses the slice as "year-0 realized-2015-interview slice mirroring the floor" — bulk shorthand from the original commit, now superseded by the per-person-anchor language in every authoritative pin (§2.8.1 body, §2.8.2, §2.8.3, §10 JSON). Purely cosmetic; worth a one-word tidy on a future touch, not a revision.

Gates

Docs-only diff (one file, +119/−37, no gate cell / threshold / floor artifact touched); 39/39 test_gate_m6_floors.py + test_gate_m6_derivations.py pass in a fresh venv; §10 m6-design-parameters JSON parses (revision 5, residual_open_decisions: none, 3 mechanical alignments); CI green on 45e757d (build, lint, pytest 3.11 + 3.13); mergeable: MERGEABLE, mergeStateStatus: CLEAN.

Disposition: VERIFIED — the amendment meets the §2.7.6 zero-choices bar; ready to merge and proceed to the harness build + engine-referee-style review + fresh registration.