gate_w1 amendment 2 (draft — ceremony pending): family-A concept cells

[MaxGhenis]

gate_w1 amendment 2 — the family-A concept cells and the C1 fingerprint (proposal, draft)

Ceremony stage: PROPOSAL (draft). This moves no threshold and edits no
gates.yaml cell — gates.yaml is byte-identical to origin/master. The
prospective flip happens in a separate ratifying PR, only after this ceremony
(proposal → adversarial referee → fixes → verification → ratify-by-merge → flip)
clears. Mirrors the amendment-1 precedent (#164 / #165).

Doc: docs/amendments/gate_w1_amendment_2_family_a_concept_cells.md.

Ceremony history

• Adversarial referee round (comment 4953763322, verdict AMEND THE AMENDMENT): confirmed the demotion boundary is principled (the referee's own five extra construction-attempts move the 65+ cells away from the frame), but caught that the blanket per-cell impossibility language was false — candidate 1 passed coresident_spouse.65+|female 4/5 — and that the retained surface oversold its own passability.
• Fixes round (comment 4953911698): rebuilt §4c with per-cell grounds (two 65+ machine reasons), added the retained-surface candor + series forecast, corrected the σ ranges; 18 → 24 bindings.
• Verification round (comment 4954135376, verdict FURTHER FIXES): every load-bearing number recomputes from the committed artifacts, and the scored_duplicate_of_demoted_married_quantity ground is ruled principled — but the fixes commit introduced one re-blanketed §2c impossibility sentence (false for the same coresident-F cell) and inherited one per-cell-false §3.1 retained-cell record, and the PR body still described the pre-fix state.
• This round (fixes 2): corrects both sentences (each now bound — the MM4/MM6 mutation analogues are now caught), folds the verifier's principled ruling + a deployment-contingent guard into §4c, reconciles the amendment-3 hh_size risk count to four (verifier note N2), and refreshes this body.

What it proposes

W1 forensics 2 (grading 4953311492, verdict "gate_w1 remains unpassable as amended — amendment 2 precedes candidate 3") established that three cell groups on the post-amendment-1 55-cell surface cannot be cleared at the gate-relevant ≥4-of-5-seed level by any contract-consistent candidate. This proposal demotes 7 cells to report-only, each with a per-cell machine reason true on the committed record:

Group	Cells	Machine reason	Why (candidate-independent)
(a) C1 fingerprint	fingerprint.ppi_nra	fingerprint_reversal_not_realized	The required PPI-over-NRA reversal does not occur under any contract-consistent transport. Forensics 1 Q5 (analytic): on the heaviest (upper-read) tail PPI savings 0.0169 ≪ NRA 0.2023 (gap 0.1854); the two candidate fingerprint runs are bit-identical (one measurement).
(b) 18-24 participation	earnings_participation.18-24|{female,male}	population_concept_delta_head_spouse_universe	A 22.1pp population-concept delta (PSID head/spouse 0.865 vs CPS all-person 0.644; male 27.6pp), not a fit-support gap. Male cell 0/5 both candidates; female caps 2/5.
(c) 65+ marital/coresident — married pair + coresident male	marital_share.married.65+|{f,m}, coresident_spouse.65+|male	cohort_vintage_hazard_frame_mismatch	Divorce over-accumulation caps the deployed married share at the CANDIDATE_16 stationary ceiling ~0.577 F / ~0.717 M; the ceiling sits below these cells' windows (males 0/5; married-65+|female caps 2–3/5). Nine permitted-lever construction-attempts each fail (§4c).
(c) 65+ — coresident female	coresident_spouse.65+|female	scored_duplicate_of_demoted_married_quantity	Candidate 1 clears this cell 4/5 — it is not unreachable. In every committed deployment coresident_spouse is regenerated as (terminal marital_state == married) (deployed ratio 1.0 every seed, bit-identical spec_resolutions), so it scores the same deployed value as the demoted married quantity against a lower window that the ~0.577 ceiling sits inside on 4/5 seeds. Demoted as a duplicate of a demoted quantity, not on impossibility.

Two machine reasons for the 65+ quad, not one — the verification round ruled the duplicate ground principled: ratio == 1.0 is a machine-checkable identity per candidate (one deployed number scored twice under two names), and the ground is deployment-contingent — a future candidate that models spousal coresidence separately (ratio ≠ 1.0) dissolves it, and re-gating is then available only via the standing forensics-then-ceremony path.

OC consequence (recomputed — differs from amendment 1)

Amendment 1 left family A byte-identical. This amendment removes 6 cells from family A itself, so the family-A faithful-candidate OC is recomputed on the residual 47-cell surface:

• Before: 53 cells → p_seed 0.922 / p_gate 0.9481 (the contract's headline; achievable OC is 0, §3).
• After: 47 cells → p_seed 0.9344 / p_gate 0.9623.

The crux (§3): the headline OC prices sampling noise, not concept/vintage bias — the 6 demoted family-A cells rate ~0.997–0.998 passable under the half-normal model, yet the forensics proves no contract-consistent candidate reaches that idealization (a 2.0σ–6.6σ systematic bias; the sub-3σ end is exactly the female cells that sometimes clear tolerance). Overall gate OC moves from an achievable 0 to 0.9623 × I(C2) on the residual 48-cell surface (47 family-A joints + the C2 fingerprint). 55 → 48 gated, 77 → 84 report-only.

Not known passable + series forecast

The amended 48-cell surface removes proven-unreachable cells; it does not establish a passable surface, and "0.9623 × I(C2)" must not be read as I(C2) being a live coin-flip. Two committed unpassability risks, forecast as amendment-3 candidates now:

• hh_size — all five cells fail 5/5 for both candidates; Q7's pre-registration was graded REFUTED (the joint roster+window lever clears only size-2, "not closable by these two entry-state levers alone"). The four cells Q7 leaves out of tolerance (sizes 1/3/4/5+) are the amendment-3 risks.
• C2's cap_150k adjacency — both candidate runs bit-identical, deployed cap_150k 16.7 years vs Smith's rank-4 1 year, τ 0.333, corrected-tail τ −1/3. C2 is retained (its failure is not yet proven candidate-independent), but the disclosure is mandatory.

Each is reopenable only via the same forensics-then-ceremony standard, never a bare demotion.

No-self-rescue (over-determined, rescue set empty)

Both committed candidates STILL FAIL on the amended 48-cell surface: candidate 1 (#162) and candidate 2 (#167) each pass 0/5 family-A seeds on the residual 47 cells and fail family C (C2 does not reverse, τ 0.333). The demotion rescues neither — over-determined via both surviving families. Per-cell truth: the demoted male cells are 0/5 for both candidates; the female cells sometimes passed (coresident-65+|female 4/5 c1), but removing a sometimes-passed cell cannot turn a 0/5 into ≥4/5 because other retained cells fail every seed.

Tests / verification

• tests/test_gate_w1_amendment2_proposal.py — 26 bindings (artifact tier) cross-check every load-bearing figure against the frozen artifacts (forensics 1/2, floors, both candidates, gates.yaml), recompute the OC on both surfaces, the per-seed no-self-rescue, and the per-cell pass records; they string-bind the §7 flip text, the three partition roll-ups, and the §8 history, and check prose assertions outside the ledger fence (de-vacuoused).
• Per-cell truth bound: candidate 1's 4/5 pass on coresident_spouse.65+|female, the two 65+ machine reasons, the ceiling-vs-window arithmetic, and the corrected 2.0–6.6σ range.
• New this round: test_section2c_no_blanket_65plus_impossibility_coresident_inside_window (MM4) pins §2c to the cube — the ~0.577 ceiling sits inside the coresident-F window on 4/5 seeds, and the false blanket cannot return; test_section3_1_5of5_fail_list_per_cell_true_excludes_never_married (MM6) pins §3.1's 5/5-fail list to the cubes and forbids re-adding marital_share.never_married.25-34|male (candidate 1 clears it 1/5). Both mutation analogues the verification round found unbound are now caught.
• Asserts the DRAFT invariant: gates.yaml still gates all 55 cells (C1 + the 6 family-A cells gate-eligible), byte-identical to master.
• Tier manifest refreshed CI-equivalent: artifact 903 → 905 (= the +2 new bindings), total 2,601 → 2,603 (tests/tier_counts.json + tests/README-tiers.md, enforced by test_tier_policy). Mutation battery re-run against the corrected sentences — the prior FM1–8 / MM1–2 set stays caught and the MM4 / MM6 analogues are now caught.

Ceremony checklist

• Proposal (this doc + this draft PR). No gates.yaml edit; no threshold moved.
• Adversarial referee round (4953763322, AMEND THE AMENDMENT).
• Fixes (4953911698).
• Verification round (4954135376, FURTHER FIXES).
• Fixes 2 (this revision: §2c + §3.1 per-cell truth + MM4/MM6 bindings; §4c deployment-contingent duplicate guard; N2 hh_size count; body refresh).
• Re-verification round.
• Ratify by merge of the flip PR (after the ceremony clears).
• Flip gates.yaml per §7 and append the §8 history entry.

Evidence chain: candidate 1 (#162), candidate 2 (#167), W1 forensics 1 (#163), W1 forensics 2 (#168); amendment 1 (#164 / #165).

🤖 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 3:58am

[MaxGhenis]

Verdict: AMEND THE AMENDMENT.

Adversarial round on the gate_w1 amendment-2 proposal (head 6e183a5), under the amendment-1 standard (4951701300 → 4951903239): every load-bearing number recomputed from the committed artifacts in a detached worktree; candidate-independence re-tested by attempting my own constructions beyond the proposal's six; the flip text attacked by mutation. Because this is the second accommodation-direction amendment to the same gate in two days, this round's central question was whether gate_w1 is being whittled to passability. Short answer (finding 1): no — the demotion boundary is principled, and I could not construct a permitted lever that rescues any demoted group at the gate-relevant ≥4-of-5 level; my own missed-lever attempts move the 65+ cells away from the frame. The blockers are narrower but real: the proposal's per-cell impossibility language is false as written for the two female 65+ cells on the committed record — candidate 1 passed coresident_spouse.65+|female on 4 of 5 seeds (finding 2) — and the ceremony record undersells two already-visible unpassability risks in the surface it retains (finding 4). The demotion set of 7 survives; the grounds and the candor do not, and this text is destined for the permanent gates.yaml history.

---

1. The whittling question, answered plainly. The boundary between demoted and retained cells tracks one principle — demote iff the committed forensics establishes that no contract-permitted lever reaches the cell; retain every cell with a live permitted lever, however badly the built candidates miss it — and the proposer applied it against self-interest:

• The 15pp concept-gap threshold was pre-registered in the forensics-2 registration (4953088871: "the 18-24 concept gap measures ≥15pp (the amendment threshold question)") before measurement; measured 22.1pp pooled (F 17.3 / M 27.6). Verified in the artifact.
• The retained surface keeps cells failing 5/5 seeds for both candidates — hh_size_share.{1..5plus}, marital_share.married.25-34|male, earnings_participation.35-44|female, marital_share.never_married.25-34|male — because each maps to a permitted lever (Q8 sex covariate: proven 4/4 with no collateral; Q6 CPS-anchored entry: adjudicated permitted, entry-component-dominant at 25-34; Q7 roster/window: permitted but see finding 4). A whittler would have demoted these; the proposal doesn't.
• The 25-34 overshoot cells are correctly not demoted: Q6's decomposition puts the miss almost entirely in the entry component (+0.092 F / +0.124 M; hazard component −0.005/+0.013), and the CPS-anchored entry lever is contract-adjudicated permitted with the right sign. The demotion boundary is the permitted-lever line, not the failure line.
• Trajectory disclosure the record should carry (see finding 4): lock 65 gated → a1 55 → a2 48; the committed evidence already names the next candidates (hh_size, C2's cap_150k adjacency). If both go the wrong way the gate lands near 43. That is not whittling-to-pass — the amended gate is still not known passable — but the record must forecast it rather than discover it one amendment at a time.

2. BLOCKER — the per-cell impossibility claims are false for the female 65+ cells on the committed record; §4c/§5/§6.3/§8 must be corrected before this text becomes history. The proposal argues the 65+ group at the male cells' magnitudes and then asserts the conclusion for the quad and for all 7 cells:

• §8 (permanent history draft): "So no contract-consistent candidate can clear any of the 7 cells." §4c: "No contract-consistent candidate can clear the 65+ quad." §5: removes "exactly the structurally-unclearable cells — no more and no fewer." §6.3: "The demoted cells were themselves failing for both candidates."
• The committed record: candidate 1 PASSED coresident_spouse.65+|female on 4 of 5 gate seeds (scores 0.129/0.164/0.142/0.156/0.173 vs tolerance 0.168; artifact pass flags [T,T,T,T,F]) — the gate-relevant clearing level. It also passed marital_share.married.65+|female on 2/5 (0.149/0.192/0.151/0.166/0.184 vs 0.163). Candidate 2: 2/5 and 1/5. The female 18-24 cell: candidate 2 passed 2/5 (0.237/0.170/0.162/0.265/0.245 vs 0.211). Per-seed pass records (c1, c2) across the six family-A demoted cells: 18-24 F 0/5, 2/5; 18-24 M 0/5, 0/5; married-65+ F 2/5, 1/5; married-65+ M 0/5, 0/5; coresident-65+ F 4/5, 2/5; coresident-65+ M 0/5, 0/5. §6.3's blanket sentence and §8's "any of the 7 cells" are contradicted by the artifacts the tests bind.
• The correct grounds, all derivable from committed numbers, are: (a) married-65+|female — the achievable deployment is capped at the CANDIDATE_16 stationary level ~0.577 (Q1: entry-0 chain lands 0.5767; Q6: entry-0.58 chain lands 0.5728 — a 40-year contraction of 0.007 per unit entry, so even an entry stock of 1.0 buys ≤ +0.004), against per-seed lower window edges [0.5619, 0.5950, 0.5798, 0.5853, 0.5904]; with ±0.01 seed-half composition the cell caps at 2–3 of 5 seeds, never 4 — unclearable at the gate level, but by ceiling-vs-window arithmetic, not by the "+0.12 divorced excess" magnitude argument (which answers the male cells' 7–8pp shortfall, not the female cell's 1.4pp). (b) coresident-65+|female — in every committed deployment coresident_spouse is regenerated as (terminal marital_state == married) (both candidates' spec_resolutions; deployed coresident/married ratio = 1.0 on all seeds), so the cell is a duplicate of the demoted married quantity scored against a slightly easier window (frame coresident/married ratio 0.973–0.992, tolerance 0.168 vs 0.163) — which is exactly why the identical deployed value straddles this window at 2–4/5 while failing the married window. Retaining it gated would (i) gate a duplicate of a demoted quantity and (ii) let a composition-blind pass certify a 65+ composition the same run gets wrong by +0.12 divorced — and since the achievable deployment sits on the pass boundary, the sampling-σ OC (0.998) wildly overprices it, the very headline-vs-achievable defect this amendment exists to cure. Demote it on those grounds. (c) the male pair and 18-24 M — the proposal's arguments hold as stated and recompute (needs 0.774–0.787/0.753–0.773 vs ceiling ~0.717; support 0.929 vs ceilings 0.750–0.814). (d) 18-24 F — support 0.8149 sits inside the pass window only on seeds 1–2 (upper edges 0.8462/0.8495; others 0.765–0.793): caps at 2/5.
• Required fix A: correct §4c/§5/§6.3/§8 to the per-cell truth; add the per-seed pass records and the ceiling-vs-window arithmetic for the female cells to the doc and the ledger, with a binding test (the fix-B standard), so the permanent record cannot carry "no candidate can clear any of the 7" alongside an artifact showing 4/5.

3. My construction attempts (the levers the proposal's table does not contain) — all fail; candidate-independence HOLDS at gate level. I implemented and ran the missed entry-timing family with the committed v2 machinery bit-for-bit (same q1 model, same CANDIDATE_16 simulate, same stream salts; K=8, full sha-verified frame):

#	Referee construction (permitted-class)	65+ married F	65+ married M	vs v2's 0.5728 / 0.7070	Verdict
R1	Per-band entry: each adult enters at their own band's lower edge, seeded from the q1 model's matching PSID entry band (65+ band: 0.5378 F / 0.6961 M)	0.4228	0.6004	worse by −0.15 / −0.11	fails 0/5
R2	Entry-55 for 65+: 65+ adults enter at 55 from the 55-64 PSID band (0.6811 F / 0.6795 M), 10–30y window	0.4951	0.6090	worse by −0.08 / −0.10	fails 0/5
R3	Max-entry bound: any entry stock ≤ 1.0 at 25, long window	≤ ~0.580	≤ ~0.717	contraction 0.007–0.010/unit (committed Q1-vs-Q6 pair)	caps married F at 2–3/5; males dead
R4	18-24 concept-matched deployment: fire the generator only for frame heads/spouses, dependents get no earnings	deployed ≈ s·0.865 < 0.52 for any head/spouse share s ≤ 0.6	—	undershoots every seed's lower edge 0.48–0.56; no permitted generator for the dependent remainder	fails
R5	C1 tail push: any career-panel treatment heavier than the upper read	—	—	the upper read is the no-zero-years maximum; sexcov/boundary/p0 all lighten the tail; PPI needs 12× (0.017→0.202)	fails

R1/R2 are the sharpest result of this round: the intuitive "seed old cohorts from old-cohort train data" lever the six-row table never tries makes the 65+ cells worse — the certified hazards' old-age dissolution erodes any short-window entry stock below the stationary level, and PSID's own 65+/55-64 entry stocks (0.538–0.681 F) sit below the frame everywhere. Together with R3, the achievable set over the entire permitted entry-state family is [0.42, 0.58] F / [0.60, 0.72] M against needs ≥0.56–0.59 F / ≥0.77–0.79 M. The 65+ demotion is not premature, and the one cell my attempts can reach (coresident F, via the stationary ceiling) is the one candidate 1 already reached — finding 2's point, not a rescue.

4. BLOCKER (candor) — the retained surface's passability is oversold by omission; two unpassability risks are already visible in committed artifacts. §5 says the residual 47 "still hold candidate-quality misses (the c3 levers …)" and calls C2 "its empirical burden"; §2a says C2 "holds under both tails and is untouched." The committed record is more specific and worse:

• hh_size (5 residual cells, all 5/5-fail for both candidates). Forensics-2 Q7's pre-registration was graded REFUTED: the two named permitted levers (coresident_parent roster + fuller fertility window) jointly clear only size-2; sizes 1/3/4/5+ improve but stay out of tolerance, and the artifact's own summary says the family "is not closable by these two entry-state levers alone" and needs "a coresidence-composition repair beyond the initial roster" — whose contract status is unadjudicated. That is a committed insufficiency finding, not a "candidate-quality miss," and §5/§9 should quote it.
• C2's full ordering has never been produced and the only measured lever moves it away. The pass rule requires the deployed order to EQUAL [elimination, +2pp, +1pp, cap_150k] with exactly the one committed swap. Both candidates deployed [elimination, cap_150k, +2pp, +1pp] (τ 0.333) with bit-identical exhaustion deltas (elim 41.0, cap_150k 16.745…, +2pp 5.705, +1pp 2.551) — the fingerprint machinery consumed the same byte-carried career panel in both runs, so the "two" C2 (and C1) empirical readings are one measurement run twice. The deployed cap_150k delay is 16.7 years against Smith's rank-4 anchor (1 year), and forensics-1's corrected (realistic) tail moves the order to [cap_150k, elim, +1pp, +2pp] — τ vs required −1/3, further away. Forensics-1's own c2_note asserts robustness only for the elim↔+2pp pair, "not the full C2 ordering." Nothing in the evidence base adjudicates whether the cap_150k adjacencies are reachable on a representative frame at all — the same question shape that demoted C1. As drafted, "0.9623 × I(C2)" invites the reading that I(C2) is a live coin flip; on every committed measurement it is 0.
• Required fix B: §5/§9 (and the §8 history) must state: (i) Q7's refuted joint-feasibility verbatim; (ii) the C2 record above, including the bit-identity of the two candidate fingerprint runs and the corrected-tail direction; (iii) the plain sentence that the amended 48-cell gate is not known passable — the amendment removes proven-impossible cells, it does not establish a passable surface; and (iv) the named next-amendment risks (hh_size 1/3/4/5+; C2's cap_150k adjacency) so amendment 3, if it comes, was forecast in this record. I considered requiring a Q5-grade robustness analysis of cap_150k in this ceremony; I am not, because C2's failure is not yet proven candidate-independent (no analytic leg exists either way, and the career-panel transport IS a live candidate lever) — but the disclosure is mandatory, and if candidate 3's fingerprint runs byte-carry the career panel again, that "empirical burden" framing expires.

5. OC and σ arithmetic: the recomputes are exact; the §3 σ ranges describe only the 18-24 pair. From the frozen floor sigmas × locked tolerances (the derivations basis): 53-cell p_seed 0.921986576938052 → 0.922, p_gate 0.9480909495222698 → 0.9481; 47-cell p_seed 0.9344024116783913 → 0.9344, p_gate 0.962342065413138 → 0.9623 — ledger-exact, and removal raises p_seed as claimed. Demoted-cell faithful pass-probs 0.996536–0.998496 → round-3 min 0.997 / max 0.998 ✓. Partition arithmetic 53/52 → 47/58, 2/0 → 1/1, 55/77 → 48/84 verified against gates.yaml. But §3's "miss the hard cells by 0.23–0.44 ln against a sampling σ of ~0.05–0.07 (3–8σ)" fits only the 18-24 cells (σ 0.0720/0.0722; scores 0.23–0.44 across both candidates). The 65+ cells miss by 0.109–0.201 ln against σ 0.0274–0.0544; the honest multiple range across the six cells and both candidates is 2.0σ–6.6σ — and the sub-3σ end is exactly the female cells that sometimes pass (finding 2). The systematic-not-stochastic crux survives (mean-of-20 draw noise is 0.002–0.007, far below every miss); the stated ranges don't. Fix with fix A.

6. No-self-rescue: clean, over-determined, recomputed per-seed. On the residual 47: candidate 1 passes 0/5 seeds (21/20/17/22/23 cells fail per seed), candidate 2 passes 0/5 (11/9/12/12/12) — the doc's 17–23 and 9–12 ranges are exact. C2 unreversed for both (τ 0.333); C1 unreversed (τ 0.667); each candidate fails both surviving families independently of the 7 demoted cells, so the rescue set is verifiably empty. The demoted male cells were 0/5 for both (rbar 0.866–0.934 vs windows; 0.701–0.729 vs 0.753–0.787). Forensics artifacts carry reported_not_gated / train_frame_side_only / publishes_regardless (verified in both), and prospective-only is stated. The one correction here is §6.3's wording (finding 2) — the conclusion stands with margin.

7. Bindings: 18/18 green and genuinely mutation-fragile. All 18 proposal tests pass in the CI-equivalent venv (0.53 s). Mutation battery (each applied to the real doc, suite rerun, tree restored; git diff clean and 18 green after):

#	Mutation	Result	Caught by
M1	delete coresident_spouse.65+|male from the §7 flip list	CAUGHT	test_section7_flip_list_binds_all_seven_demoted_cells
M2	§7 overall roll-up 48/84 → 49/83	CAUGHT	test_section7_three_rollups_match_ledger_partition
M3	§8 history p_gate 0.9623 → 0.9624	CAUGHT	test_section8_history_figures_bound
M4	strip the 65+ machine reason from every prose occurrence (ledger fence intact)	CAUGHT	test_machine_reasons_named_in_prose_outside_ledger
M5 (mine)	§7 OC transition 0.9623 → 0.9632	CAUGHT	test_section7_oc_and_cube_transitions_bound
M6 (mine)	§8 history concept gap 22.1pp → 21.1pp	CAUGHT	test_section8_history_figures_bound

The de-vacuousing holds (the ledger strip is proven before the prose assertions run). Fix A/B should extend the ledger + bindings to the corrected per-cell claims.

8. Hygiene: clean. gates.yaml byte-identical to origin/master (cmp clean) — the draft moves nothing; all 55 cells still gated (C1 + the 6 family-A cells gate-eligible, verified via the suite and directly). Diff touches exactly the 4 claimed files. Tier manifest live-verified in the CI-equivalent venv: unit 225, artifact 897, integration_psid 797, reproduction_legacy 520, oracle_policyengine 156 — total 2,595, every tier exact; artifact 879 → 897 = +18 = precisely the new binding suite, and the README table (which had drifted from the enforced manifest pre-PR: 216/854/790) is now consistent. Fresh CI green on head 6e183a5 (build, lint, pytest 3.11 + 3.13; deploy skipping is the draft norm); MERGEABLE / CLEAN; the PR remains a draft. Registration/grading pointers verified (4953088871 in-artifact; 4953311492 is the forensics-2 grading; note the artifact's own grading_pointer 4953064479 is candidate-2's grading — the ledger's pointer is right, a reader tracing the artifact should not be confused by the two).

9. Text corrections (non-blocking beyond fixes A/B). (i) §4a's anchor sentence is garbled: "…the Mermin Table 1 payroll-savings ordering, in which PPI (−0.14) already ranks below NRA→70 (−0.50) is false — Mermin has PPI (−0.14) above NRA (−0.50) in signed savings…" — two drafts spliced; the intended (correct) content is that Mermin's signed order has PPI above NRA while the required magnitude reversal is the falsified hypothesis. Rewrite. (ii) §1/§2a/§8's "two empirical candidate confirmations" for C1 should note the two runs are bit-identical (one computation; the analytic Q5 leg is what carries the demotion — it does). (iii) v2 module docstring says the entry model "lands 0.467" at 25-34F where the committed q1 model value is 0.577 — stale figure in a non-amendment file; flag for the flip PR's hygiene sweep, no action here.

---

Required before verification round:

• Fix A (per-cell truth, finding 2 + 5): correct §4c/§5/§6.3/§8; add per-seed pass records for the 6 family-A demoted cells and the female-cell ceiling-vs-window arithmetic; ground the coresident-F demotion in the duplicate-quantity/boundary-straddling/composition-blind argument, not "no candidate lever exists"; fix §3's σ ranges (2.0–6.6σ; 0.11–0.44 ln; σ 0.027–0.072); bind the new figures.
• Fix B (retained-surface candor, finding 4): Q7-refuted verbatim; the full C2 record (bit-identical candidate runs, cap_150k 16.7 vs Smith rank-4, corrected-tail τ −1/3, pair-only robustness); the "not known passable" sentence; the named amendment-3 risks — in §5/§9 and the §8 history.
• Fix C (text, finding 9i–ii): repair §4a; qualify the C1 empirical legs.

The demotion set of 7 is ratifiable once the record says true things about each of the 7. The construction-attempt method was applied and the amendment survives it — including two levers its own table missed — which is precisely what the record should be able to say, in the proposer's corrected words rather than the referee's.

[MaxGhenis]

Fixes round — AMEND THE AMENDMENT addressed (head 8e713ef)

Thank you for the round. The whittling question mattered most and the answer held; the two blockers were real and are fixed. The demotion set of 7 is unchanged, gates.yaml stays byte-identical to origin/master, and the OC numbers are untouched (you reproduced them exact). Each finding → fix → where it is bound:

#	Finding	Fix	Bound by
F1	Per-cell impossibility false: candidate 1 passed coresident_spouse.65+|female 4/5	§4c rebuilt with per-cell grounds: the married pair + coresident_spouse.65+|male carry cohort_vintage_hazard_frame_mismatch (stationary ceiling ~0.577 F / ~0.717 M the deployment cannot reach; married-65+|female caps 2–3/5 by ceiling-vs-window, not the +0.12 magnitude); coresident_spouse.65+|female carries a new reason scored_duplicate_of_demoted_married_quantity (deployed coresident/married ratio 1.0 on every seed → duplicate of the demoted married quantity, straddling a slightly easier window; the 0.998 sampling-σ OC overprices a boundary cell). The 4/5 pass is stated plainly in §1/§4c/§6/§8.	test_per_cell_pass_records_bound_to_committed_artifacts (recomputes every per-seed count from the cubes), test_65plus_two_machine_reasons_grounds_are_per_cell_true, test_ceiling_vs_window_65plus_female_arithmetic_bound (Q1 0.5767 / Q6 0.5728 / window edges / ratio 1.0), test_section8_history_figures_bound (asserts the false blanket sentence is absent)
F2	Retained surface's passability oversold; Q7 REFUTED + C2 record omitted	§5/§9 + §8 now state the 48-cell surface is NOT KNOWN PASSABLE; Q7's REFUTED verbatim ("not closable by these two entry-state levers alone"; all 5 hh_size cells 0/5 both candidates); the full C2 record (bit-identical runs, cap_150k 16.7y vs Smith rank-4 1y, corrected-tail τ −1/3, pair-only robustness). Series forecast added: amendment-3 risks (5 hh_size + C2 cap_150k) pre-named, each reopenable only via the same forensics-then-ceremony standard.	test_retained_surface_not_known_passable_candor, test_c2_record_bound_and_amendment3_forecast (recomputes the corrected-tail τ = −0.3333 and cap_150k 16.7 from the artifacts)
F3	§3 "3–8σ" describes only the 18-24 pair	Corrected to the true per-cell 2.0σ–6.6σ (0.11–0.44 ln vs σ 0.027–0.072); the sub-3σ end is named as the female cells that sometimes pass	test_sigma_range_corrected_2_to_6p6 (recomputes the range from candidate scores × floor σ; asserts "3–8σ" absent)
F4	Table missing the referee's constructions	§4c incorporates R1 per-band entry (0.4228/0.6004), R2 entry-55 (0.4951/0.6090), R3 max-entry bound, citing comment 4953763322; the achievable-set bound [0.42,0.58]F / [0.60,0.72]M vs needs ≥0.56–0.59F / ≥0.77–0.79M is stated	test_q6_contract_adjudication_says_65plus_unfixable_by_permitted_lever (asserts R1/R2 landed values + the comment id in §4c)
F9	§4a Mermin sentence garbled; C1 legs	§4a rewritten (Mermin signed order has PPI above NRA; the falsified hypothesis is the magnitude reversal); C1's two candidate legs qualified as one bit-identical measurement, the analytic Q5 leg carrying the demotion	test_c1_non_reversal_bound_and_runs_bit_identical (asserts the two runs' fingerprints are byte-equal)

Verification standard held. Bindings 24/24 green in the CI-equivalent venv (was 18; +6 for the per-cell/ceiling/σ/C2/candor evidence). Mutation battery re-run against the corrected sentences — all caught: reintroducing "no candidate can clear any of the 7" in §8, drifting the coresident-F pass 4→3, mislabeling the coresident-F reason, σ 2.0→3.0, cap_150k 16.7→15.0, the §7 roll-up, a ceiling window-edge digit, and a full removal of the "not known passable" candor. Tier manifest recounted CI-equivalent (artifact 897 → 903 = the +6 new bindings; tier_counts.json + README-tiers.md). gates.yaml byte-identical to master; the draft moves nothing.

Ready for the verification round.

[MaxGhenis]

Verification round — gate_w1 amendment 2 (head 8e713ef)

Verdict: FURTHER FIXES — two per-cell-truth sentences and one record-hygiene item, all text-tier. Every load-bearing number in the fixed state (head 8e713ef) recomputes exactly from the committed artifacts in a detached worktree with a disposable env; the demotion set of 7, the two-reason 65+ ledger, the candor block, the series forecast, the OC transition, no-self-rescue, and all 24 bindings verify clean, and the fixes lane's mutation battery reproduces. The two blockers from 4953763322 are genuinely fixed where the referee pointed. But the fixes commit introduced one new quad-level impossibility sentence that is false for coresident_spouse.65+|female on the committed record — the exact cell and the exact defect class F1 was about — and §3.1 carries a per-cell-false retained-surface record inherited verbatim from the adversarial round itself. Because this text is one merge from the permanent record, and because the standard this ceremony set is "the record says true things about each cell," those two sentences must be corrected (and bound) before ratification. Both are one-line edits; nothing structural moves.

---

1. Per-cell truth of every ground — VERIFIED, with two exceptions (fixes 1–2 below). Recomputed per-seed from the candidate cubes (runs/gate_w1_candidate1_v1.json, gate_w1_candidate2_v1.json):

Cell	c1	c2	Ledger/doc
earnings_participation.18-24|female	0/5	2/5	✓
earnings_participation.18-24|male	0/5	0/5	✓
marital_share.married.65+|female	2/5 [T,F,T,F,F]	1/5	✓
marital_share.married.65+|male	0/5	0/5	✓
coresident_spouse.65+|female	4/5 [T,T,T,T,F], scores 0.1289/0.1645/0.1420/0.1563/0.1731 vs tol 0.168	2/5	✓ — stated plainly in §1/§4c/§6/§8
coresident_spouse.65+|male	0/5	0/5	✓

• Stationary-ceiling arithmetic: Q1 entry-0 65+F = 0.5766873… → 0.5767 ✓; Q6 entry-0.58 = 0.5727633… → 0.5728 ✓; implied |slope| 0.0068/unit ≈ "~0.007" ✓; Q1 65+M = 0.7123 (the doc's "~0.717" is the referee's R3 max-entry bound; either reading leaves the male cells ≥4.5pp short of their windows). Married-F per-seed lower edges recompute to [0.5619, 0.5950, 0.5798, 0.5853, 0.5904] ✓ (= ledger); c1 deployed rbar 0.5696–0.5867 ("0.570–0.587" ✓) clears exactly the two edges below the ceiling → 2/5 ✓, c2 1/5 ✓; "caps 2–3/5, never 4" is consistent with the committed records and the edge geometry (three edges ≥ 0.5853). Male windows: married needs 0.7739–0.7869 ("≥0.774–0.787" ✓), coresident 0.7528–0.7727 ("≥0.753–0.773" ✓), 0/5 both candidates ✓.
• Duplication claim: deployed coresident/married ratio = 1.000000 on every seed, both candidates, both sexes ✓; both artifacts' spec_resolutions state coresident_spouse is RE-GENERATED as (terminal marital_state == married) ✓; frame coresident/married ratio 0.9726–0.9918 ("~0.97–0.99" ✓); tolerance 0.168 vs 0.163 ✓ — the identical deployed value straddling the easier window at 4/5 while failing the married window at 2/5 is exactly what the cubes show.
• 18-24: concept gap 22.1pp pooled / 17.3 F / 27.6 M, PSID 0.865 vs CPS 0.644, threshold flag true ✓; F support 0.8149 admits only the two loosest upper edges (0.8462/0.8495; others 0.7654–0.7929 = "0.765–0.793") ✓; M support 0.9287 vs ceilings 0.7496–0.8140 ("0.750–0.814") ✓. Divorce excess +0.120/+0.127, deployed/frame married 0.573/0.691 and 0.707/0.856, widowhood not realized ✓.
• Blanket-impossibility hunt: §8's banned sentence is gone and the binding forbids its return ✓; §1's and §3's group-level statements are true at group level and §3 explicitly disambiguates ("the group is not 'no candidate can clear any cell'") ✓; §5's "no more and no fewer" now carries the plus-clause and is true as a set identity ✓. Two per-cell-false survivors found — one blanket-impossibility rephrasing (§2c, fix 1) and one false retained-cell record (§3.1, fix 2).

2. Principle coherence — scored_duplicate_of_demoted_married_quantity is PRINCIPLED. Ruled explicitly: this is quantity-level deduplication, not demotion of an inconvenient passing cell, on four grounds. (a) The duplication is identity, not correlation: ratio exactly 1.0 on every committed seed for every committed candidate, because every committed deployment regenerates coresident as (married) — the gate was scoring one deployed number twice under two names, so demoting the married quantity while gating its alias would be a nominal demotion; the boundary principle survives restated at the quantity level (demote iff no permitted lever reaches the quantity; a bit-identical alias is the same quantity under another name). (b) Single mechanism, single disposition: the mechanism adjudicated unreachable (stationary ceiling vs frame window) is the same behind both cells; the coresident window is merely translated (frame ratio 0.97–0.99, tol 0.168 vs 0.163) so it happens to straddle the ceiling — keeping it gated would re-certify under the loose window what the tight window was just adjudicated to un-certify. (c) The forfeited pass is a false certificate: it would certify a 65+ composition the same run gets +0.120 divorced wrong, and its 0.998 half-normal OC is maximally mispriced for a boundary-straddling cell — the headline-vs-achievable defect this amendment exists to cure. (d) It cannot whittle: both candidates remain 0/5 on the residual 47 (recomputed below), so the demotion rescues nothing; and the ratio==1.0 condition is machine-checkable per candidate, which keeps the rider narrow — merely-correlated cells cannot be swept in under this reason. One forward-looking guard, non-blocking: the ground is deployment-contingent — if a future candidate models spousal coresidence separately (ratio ≠ 1.0), the duplicate ground dissolves for that world; re-gating is then available only via the standing ceremony path, which report-only + §10 already enforce. State it at flip time if convenient.

3. Candor bindings — VERIFIED. "Not known passable" present in §1/§5/§9 prose + §8 history (NOT KNOWN PASSABLE) ✓. Q7: pre_registration_proves_3_4_5plus false, joint construction clears only size-2 (score 0.1159 vs tol 0.123; sizes 1/3/4/5+ out), all five hh_size cells 0/5 for both candidates ✓, and both quotes are verbatim in gate_w1_forensics2_v1.json ("not closable by these two entry-state levers alone"; "a coresidence-composition repair beyond the initial roster") ✓. C2 record: fingerprint blocks bit-identical across the two candidate artifacts ✓ (C1 block too ✓); deltas elim 41.0 / cap_150k 16.745 / +2pp 5.705 / +1pp 2.551 ✓; Smith rank-4 cap_150k = 1 year ✓; corrected-tail order [cap_150k, elim, +1pp, +2pp] gives τ = (2−4)/6 = −1/3 vs required ✓; c2_note verbatim: "only the committed elimination<->+2pp pair is asserted robust, not the full C2 ordering" ✓. Series forecast pre-names the hh_size cells + C2 cap_150k, reopenable only via forensics-then-ceremony, with the C2 expiry condition (byte-carried panel in candidate 3) ✓. One harmonization note (N2) on "five" vs the ledger's four.

4. Bindings + mutations. 24/24 green in 0.52–0.57 s in a disposable env (fresh uv venv + editable install — the worktree/shared-venv trap avoided; test_contract_identity 8/8 green there too), and the +6 delta vs the proposal commit is exactly the six claimed tests (plus the telling rename of test_no_self_rescue_demoted_cells_were_failing_for_both_candidates to the corrected-wording binding). CI-equivalence: the CI invocation is plain pytest -q, and CI itself ran it green on this exact head (item 6); locally I additionally ran the full unit tier (223 passed / 2 skipped, 2,376 deselected = the full 2,601 collected), whose test_tier_policy validates the manifest against the complete pre-deselection inventory. Mutation battery — the fixes lane's eight re-run plus six of my own, each applied to the real doc, suite rerun, file restored, git diff clean and 24 green after:

#	Mutation	Result
FM1	reintroduce "no contract-consistent candidate can clear any of the 7 cells" in §8	CAUGHT — test_section8_history_figures_bound
FM2	ledger coresident-F pass 4→3	CAUGHT — test_per_cell_pass_records… + test_65plus_two_machine_reasons…
FM3	ledger coresident-F reason mislabeled to ceiling	CAUGHT
FM4	σ multiple 2.0→3.0	CAUGHT
FM5	cap_150k 16.7→15.0	CAUGHT
FM6	§7 roll-up 48/84→49/83	CAUGHT
FM7	window-edge digit 0.5619→0.5629	CAUGHT
FM8	full removal of "not known passable"	CAUGHT
MM1	§7 flip list: coresident-F reason flipped	CAUGHT — test_section7_flip_list…and_reasons
MM2	ledger drops c2_cap_150k_adjacency from amendment-3 risks	CAUGHT
MM3	§4c prose-only ceiling tweak 0.5767→0.5776 (ledger intact)	MISSED
MM4	§2c quad sentence strengthened to the original blocker's form	MISSED
MM5	§1 "clears on 4 of 5"→"3 of 5" (§4c intact)	MISSED
MM6	§3.1 "failing 5/5 for both candidates"→"4/5"	MISSED

MM3/MM5 are the known, accepted limits of string-binding (the ledger is the bound record, and any one surviving occurrence satisfies a prose-presence assert — dropping only §5's "not known passable" instance is likewise tolerated while the full removal is caught) — no action. MM4/MM6 are not: they show the two false sentences below sit exactly where no binding reaches.

5. Invariants — ALL VERIFIED. gates.yaml byte-identical to origin/master (cmp clean); all 55 cells still gated; the diff touches exactly the 4 claimed files. OC: 53-cell p_seed 0.9219865769381 / p_gate 0.9480909495223 → 0.922/0.9481 = the floors headline ✓; 47-cell p_seed 0.9344024116784 / p_gate 0.9623420654131 → 0.9344/0.9623 ✓ (own half-normal implementation from frozen floor σ × locked tolerances); demoted passprobs 0.996536–0.998496 → 0.997/0.998 ✓. §7 flip text string-bound (all 7 cells, all 4 reasons, the coresident-F→duplicate mapping regex-bound); roll-ups 47/58, 1/1, 48/84 recompute from gates.yaml (52+25+0 = 77 report-only now ✓). No-self-rescue: c1 passes 0/5 on the residual 47 (per-seed fails 21/20/17/22/23), c2 0/5 (11/9/12/12/12), C2 unreversed both (τ 0.3333), C1 unreversed (τ 0.6667) — over-determined, rescue set verifiably empty ✓. §8 history draft: figures bound and recomputed; placeholder pattern matches the amendment-1 precedent; evidence chain (#162/#167/#163/#168 + 4953763322) correct; the entry is correctly NOT yet in gates.yaml. σ: ln misses 0.1092–0.4358 → 0.11–0.44 ✓, σ 0.0274–0.0722 → 0.027–0.072 ✓, per-seed multiple min 2.01 → 2.0 ✓; the stated 6.6 upper sits between the mean-of-seeds max (6.54) and the per-seed max (7.27) — the binding asserts exactly that bracketing, and the stated range understates the worst per-seed multiple (conservative direction; N3). Draw noise recomputes to 0.0024–0.0071 ("0.002–0.007") ✓. Tier manifest live-verified CI-equivalently: unit 225 / artifact 903 / integration_psid 797 / reproduction_legacy 520 / oracle_policyengine 156, total 2,601 — every tier exact; 897→903 = precisely the +6 new bindings; README table consistent.

6. Fresh CI on 8e713ef: mergeable: MERGEABLE, mergeStateStatus: CLEAN; gh pr checks 169 exit 0 — build, lint, pytest 3.11 (3m9s), pytest 3.13 (3m5s) all pass; both workflow runs confirmed on head_sha 8e713ef… via the API; deploy "skipping" is the draft norm. The PR remains a draft.

---

Required fixes (all text-tier; the demotion set, ledger, tests, and arithmetic need no change):

1. §2c (introduced by the fixes commit — not present at 6e183a5): "These ceilings sit below the frame's pass windows (§4c), which is why the 65+ cells cannot be cleared by any permitted entry-state lever." False for coresident_spouse.65+|female on the committed record — candidate 1 cleared it 4/5 under a permitted deployment, and the ceiling does not sit below that cell's window: the coresident-F lower edges are [0.5478, 0.5759, 0.5717, 0.5769, 0.5813], so the ~0.577–0.581 ceiling sits inside the window on 4 of 5 seeds — which is §4c's own story. This is the F1 defect class re-instantiated for the same cell, one section earlier. Rescope to the per-cell truth (e.g., "…which is why the married pair and coresident_spouse.65+|male cannot be cleared by any permitted entry-state lever, married-65+|female caps at 2–3/5, and the coresident-65+|female window straddles the ceiling — §4c") and extend a binding to pin §2c (MM4 shows the sentence is currently unreachable by the suite).
2. §3.1: "the retained surface keeps cells failing 5/5 for both candidates — hh_size_share.{1..5plus}, marital_share.married.25-34|male, earnings_participation.35-44|female, marital_share.never_married.25-34|male" — false for the last cell: candidate 1 passed marital_share.never_married.25-34|male on seed 1 (score 0.1707 vs tol 0.192; flags [T,F,F,F,F] = 4/5-fail; candidate 2 is 0/5). The sentence was inherited verbatim from finding 1 of the adversarial round (4953763322) — the round's own retained-surface record carried a per-cell inaccuracy, and this ceremony's standard applies to it too. The whittling conclusion is untouched (a 1/5-passing cell retained gated is still retention against self-interest; its Q6 CPS-anchored entry lever is still live). Correct to the per-cell record and bind it (MM6 shows §3.1 is unbound). The other three "5/5" claims (§5/§8/§9 — hh_size-only) recompute true and need no change.
3. PR body refresh before ratify-by-merge. The body still describes the pre-fix state: one machine reason for the whole 65+ quad (no scored_duplicate_of_demoted_married_quantity), "18 bindings", "artifact 879 → 897", and none of the mandated candor (no per-cell pass records, no "not known passable", no series forecast). Under ratify-by-merge the body becomes the permanent merge-commit message; amendment 1's fixes round included exactly this "PR-body precision corrections" step. Bring the body to the fixed state (two-reason table, coresident-F 4/5 stated plainly, 24 bindings, artifact 903, candor + forecast).

Non-blocking notes: N1 — §5's parenthetical enumerates 5 of the 6 unreachable cells (omits earnings_participation.18-24|female, whose ≥4/5-unreachability §4b establishes: the 0.8149 support fits only the two loosest windows, 0.846/0.850, exceeding the other three upper edges 0.765–0.793 → caps 2/5); the set identity is true, the enumeration is short one example — tidy with fix 1 or at flip. N2 — prose says "the five hh_size cells" as amendment-3 risks while the bound ledger (and fix B's own wording) names four (1/3/4/5plus; size-2 is the one cell Q7's joint levers clear); conservative direction, and §8's parenthetical facts are true of all five — harmonize prose vs ledger at flip time. N3 — σ 6.6 upper: see item 5; no action. N4 — earnings_participation.35-44|female per-seed scores are bit-identical between the two candidate runs (0.1168/0.1388/0.0965/0.1000/0.0994), i.e. the interior earnings machinery was byte-carried like the fingerprint panel; harmless here, but the same candidate-3 expiry logic the doc applies to C2 would apply if a third identical vector appears.

The corrected grounds are true, the boundary principle survives its own exception (ruled principled above), the candor is real and bound, and every number in the record is the number in the artifacts. Fix the two sentences, refresh the body, and this is ratifiable — the verification round should then have nothing left to say, which is the point.

[MaxGhenis]

Fixes 2 — verification round (4954135376, FURTHER FIXES) addressed (head dd913ac)

Thank you for the verification round. Every load-bearing number was confirmed to recompute from the committed artifacts, and the scored_duplicate_of_demoted_married_quantity ground was ruled principled — so nothing structural moved. The three text-tier items and note N2 are fixed, each now bound, and the two mutation analogues the round found unbound (MM4, MM6) are now caught.

#	Item (verification round)	Fix	New / extended binding
1	§2c blanket — the fixes commit re-instantiated the F1 blanket ("the 65+ cells cannot be cleared by any permitted entry-state lever"), false for coresident_spouse.65+|female	Rewritten to per-cell truth: the ceiling-below-window relation holds for the married pair + coresident male; for coresident female the ~0.577 ceiling sits inside the (lower) window on 4 of 5 seeds (lower edges [0.5478, 0.5759, 0.5717, 0.5769, 0.5813]) — which is why candidate 1 clears it 4/5 and its ground is duplication, not unreachability	test_section2c_no_blanket_65plus_impossibility_coresident_inside_window — recomputes from candidate 1's cube that the deployed value (== married rbar, ratio 1.0) is inside the coresident-F window on exactly 4/5, forbids the blanket + its paraphrases, requires the per-cell correction. MM4 now CAUGHT.
2	§3.1 false 5/5 — marital_share.never_married.25-34|male listed as failing 5/5 for both candidates (candidate 1 passes seed 1, 0.1707 vs tol 0.192, [T,F,F,F,F])	Removed from the 5/5-fail list; the complete gate-eligible 5/5-fail retained set is hh_size_share.{1..5plus} + marital_share.married.25-34|male + earnings_participation.35-44|female (verified from both cubes). never_married.25-34|male is retained a fortiori (a cell candidate 1 clears 1/5, retained against self-interest), its Q6 lever still live	test_section3_1_5of5_fail_list_per_cell_true_excludes_never_married — parses the §3.1 5/5-fail clause, forbids never_married.25-34|male in it, recomputes the genuine 5/5-fail set from both cubes, pins the "5/5" count. MM6 (both the reintroduce-membership and the 5/5→4/5 count variants) now CAUGHT.
3	PR body stale (single 65+ reason, "18 bindings", "879→897", no candor)	Body rewritten via gh pr edit --body-file: two-machine-reason table, coresident-F 4/5 stated plainly, 26 bindings, artifact tier 903→905, not-known-passable candor + series forecast, and the full ceremony history (adversarial 4953763322 → fixes 4953911698 → verification 4954135376 → this round)	n/a (permanent squash message)
4	Principled ruling into §4c's duplicate ground	Added the verifier's ruling + deployment-contingent guard: ratio == 1.0 is a machine-checkable identity per candidate (both deployments' spec_resolutions regenerate coresident_spouse as (terminal marital_state == married), bit-identical); the ground dissolves for a future candidate that models coresidence separately (ratio ≠ 1.0), and re-gating then requires the standing forensics-then-ceremony path — citing comment 4954135376	test_65plus_two_machine_reasons_... + test_ceiling_vs_window_... (ratio 1.0 per seed both candidates)
5	N2 — prose "five" vs ledger "four" hh_size amendment-3 risks	Reconciled to four (sizes 1/3/4/5+) in §1/§5/§8/§9: Q7's joint lever clears only size-2 (0.1159 vs tol 0.123), so only these four are at risk. All five hh_size cells still fail 5/5 (the fact is retained)	Folded into test_c2_record_bound_and_amendment3_forecast — the ledger names exactly the four, Q7 clears size-2 (recomputed), prose names the four sizes

Mutation battery (each applied to the real doc, suite rerun, file restored, tree clean + 26 green after):

#	Mutation	Result
FM1	reintroduce "clear any of the 7" in §8	CAUGHT
FM2	ledger coresident-F pass 4→3	CAUGHT
FM3	ledger coresident-F reason mislabeled to ceiling	CAUGHT
FM4	σ multiple 2.0→3.0	CAUGHT
FM5	cap_150k 16.7→15.0	CAUGHT
FM6	§7 roll-up 48/84→49/83	CAUGHT
FM7	window-edge digit 0.5619→0.5629	CAUGHT
FM8	full removal of "not known passable" (all case variants)	CAUGHT
MM1	§7 flip list: coresident-F reason flipped	CAUGHT
MM2	ledger drops c2_cap_150k_adjacency	CAUGHT
MM4	§2c re-blanketed to the original blocker form	CAUGHT (was MISSED)
MM6a	§3.1 reintroduce never_married.25-34|male into the 5/5-fail list	CAUGHT (was MISSED)
MM6b	§3.1 count word 5/5 → 4/5	CAUGHT (was MISSED)

(MM3/MM5 — prose-only ceiling tweak with the ledger intact, and a single-instance "4 of 5"→"3 of 5" with §4c intact — remain the accepted limits of string-binding, per the verification round; no action.)

Bindings 24 → 26. Tier manifest refreshed CI-equivalent: artifact 903 → 905 (= the +2 new bindings), total 2,601 → 2,603 (tests/tier_counts.json + tests/README-tiers.md, enforced by test_tier_policy against the full pre-deselection inventory). The full always-runnable suite (unit + artifact) is green in a disposable env (1,089 passed / 41 skipped / 1,473 deselected; test_contract_identity green there). black -l 79 + ruff clean.

gates.yaml byte-identical to origin/master; the draft moves nothing. The PR remains a draft — ready for the re-verification round.

[MaxGhenis]

Re-verification round — gate_w1 amendment 2 fixes 2 (head dd913ac)

Verdict: RATIFY AS-IS. All three verification-round items were mine, so I checked resolution, not renaming — recomputing the corrected sentences from the cubes and re-running the mutation analogues the round found unbound. Each is genuinely fixed and now bound; the two extras (the §4c principled-ruling fold and the N2 count reconciliation) landed faithfully; every prior invariant holds byte-for-byte on the new head. The diff 8e713ef→dd913ac touches only the amendment doc, the bindings test, tier_counts.json, and README-tiers.md — no artifact, no gates.yaml, no scope creep. Nothing remains for a further round. Ratification is yours.

---

Item 1 — §2c per-cell corrected. RESOLVED + BOUND. §2c now opens "the ceiling-below-window relation holds for three of the four, not the quad," scopes "cannot be cleared by any permitted entry-state lever" to the married pair + coresident_spouse.65+|male, and carves out coresident_spouse.65+|female explicitly: the ~0.577 ceiling sits inside that cell's window on 4 of 5 seeds, per-seed lower edges [0.5478, 0.5759, 0.5717, 0.5769, 0.5813] — which I recompute exactly from candidate 1's cube (rate_a·e^−tol), with rbar ≥ lower_edge on seeds 1–4 and the cell clearing 4/5 (flags [T,T,T,T,F]). The false blanket is gone; the whole-doc sweep finds no surviving per-cell impossibility claim (the only "cannot be cleared" hits are the group-scoped §1 framing — true at the ≥4/5 group level — the corrected §2c, §3's explicit disambiguation, and the fixes-round text quoting the banned phrase to mark it false). New binding test_section2c_no_blanket_65plus_impossibility_coresident_inside_window present. MM4 re-run: now CAUGHT — reintroducing "65+ cells cannot be cleared by any permitted entry-state lever" and stripping the "inside … 4 of 5" correction both fail the new test (each 1 failed, 25 passed).

Item 2 — §3.1 5/5-fail list corrected. RESOLVED + BOUND, enumeration recomputed independently. I recomputed the genuine 5/5-fail-both retained (gated) set directly from both cubes — cells on the residual 47 surface with _seed_pass_counts == 0 for both candidates — and got exactly seven: hh_size_share.{1,2,3,4,5plus} + marital_share.married.25-34|male + earnings_participation.35-44|female, which is precisely §3.1's corrected enumeration (hh_size×5 + the two concretes). marital_share.never_married.25-34|male is correctly excluded: candidate 1 clears it on seed 1 (score 0.1707 vs tol 0.192; flags [T,F,F,F,F] = 1/5), candidate 2 0/5 — and §3.1 now states that per-cell record and keeps the cell gated "against self-interest a fortiori." New binding test_section3_1_5of5_fail_list_per_cell_true_excludes_never_married present. MM6 re-run: both CAUGHT — MM6a (reintroduce never_married.25-34|male into the complete-set clause) and MM6b (downgrade "5/5 for both candidates" → "4/5") each fail the new test.

Item 3 — PR body current. RESOLVED. The body is now the fixed-state squash message: the two-reason 65+ table (row (c) married-pair+male → cohort_vintage_hazard_frame_mismatch; row (c) female → scored_duplicate_of_demoted_married_quantity), coresident-F "Candidate 1 clears this cell 4/5" stated plainly, 26 bindings, artifact 903 → 905, the "Not known passable + series forecast" candor block (hh_size four risks, C2 cap_150k), and the full Ceremony history through fixes 2. No pre-fix residue.

Extra A — §4c principled ruling + deployment-contingent guard. FAITHFUL. The folded passage attributes to verification round 4954135376 and renders my ruling without distortion: identity-not-correlation ("ratio == 1.0 … machine-checkable identity per candidate, not a correlation that a tighter candidate might loosen"), the "one deployed number twice under two names" framing, the verbatim quantity-level restatement ("demote iff no permitted lever reaches the quantity; a bit-identical alias is the same quantity under another name"), the machine-enforced narrowness (merely-correlated cells can't be swept in), and the deployment-contingent guard (a future ratio ≠ 1.0 candidate dissolves the ground; re-gating only via forensics-then-ceremony). This is my reasoning, correctly quoted.

Extra B — amendment-3 hh_size count reconciled to four. CORRECT, size-2 number recomputed. §5/§9/§8/§1 now say the four cells Q7 leaves out of tolerance (sizes 1/3/4/5+) are the amendment-3 risks, while keeping the candor that all five hh_size cells fail 5/5 in the deployed candidates. I recomputed Q7's joint-lever scoring from gate_w1_forensics2_v1.json: size-2 score 0.1159 vs tol 0.123 → clears, sizes 1/3/4/5+ all out (0.8251/0.2727/0.2444/0.2822 vs 0.191/0.191/0.174/0.184) — so exactly one of the five is reachable by Q7's best lever, and four are standing risks. The distinction (deployed-candidate 5/5-fail vs forensic-lever reachability) is exactly right. Ledger amendment3_risks now holds the four hh_size + c2_cap_150k_adjacency; N2a mutation (re-add hh_size_share.2) CAUGHT by test_c2_record_bound_and_amendment3_forecast.

Invariants — all re-verified on dd913ac. 26/26 bindings green in a disposable env (fresh uv venv + editable install; 0.85 s). gates.yaml byte-identical to origin/master (cmp clean); all 55 cells still gated; the origin/master…HEAD diff is exactly {amendment doc (A), bindings test (A), tier_counts.json (M), README-tiers.md (M)}. Tier manifest live-verified CI-equivalently: unit 225 / artifact 905 / integration_psid 797 / reproduction_legacy 520 / oracle_policyengine 156 = 2,603; 903→905 = precisely the +2 new bindings; README consistent. The unchanged load-bearing numbers still recompute (OC 0.9344/0.9623, no-self-rescue 0/5 both with fails 21/20/17/22/23 and 11/9/12/12/12, C2 τ 0.3333 / corrected −1/3, σ 2.0–6.6). Fresh CI on dd913ac: MERGEABLE, gh pr checks 169 exit 0 — build, lint, pytest 3.11 (3m12s), pytest 3.13 (3m23s) pass, runs confirmed on head_sha dd913ac…; deploy "skipping" is the draft norm; PR remains a draft.

Two non-blocking string-binding limits (same class as the verification round's accepted MM3/MM5, noted for completeness — no action): flipping §3.1's qualitative verb "clears"→"fails" for the never_married record is not caught (the load-bearing "1/5" and "single loosest seed" tokens are pinned, but the verb is not — a self-contradictory sentence, not a false number); and changing one redundant prose occurrence of "1/3/4/5+"→"1/2/3/4/5+" is not caught (the presence-assert is satisfied by another occurrence — but the ledger drift is caught by N2a). Both leave every load-bearing figure bound.

The record now says true things about each of the seven cells and about the retained surface, the boundary principle survives its own exception, and every number in the doc is the number in the artifacts. Ratifiable as-is.