gate_w1 amendment 2 flip: demote family-A concept cells + C1 to report-only (ratified c50f8bb)

[MaxGhenis]

Applies the RATIFIED gate_w1 amendment 2 — 2026-07-12-w1-family-a-concept-cells, ratified by merge of PR #169 (squash c50f8bb) — to gates.yaml, per the string-bound §7 flip text of docs/amendments/gate_w1_amendment_2_family_a_concept_cells.md.

§7 → yaml application table

§7 item	Path	Applied
1	thresholds.family_a views → report_only + report_reasons + retained_tolerances	6 cells moved; per-cell machine reasons (2× population_concept_delta_head_spouse_universe, 3× cohort_vintage_hazard_frame_mismatch, 1× scored_duplicate_of_demoted_married_quantity); tolerances retained verbatim (0.211/0.221/0.163/0.084/0.168/0.094)
1	family_a.faithful_candidate_oc	53→47 cells, p_seed 0.922→0.9344, p_gate 0.9481→0.9623
2	protocol.fresh_run_artifact_schema.per_draw_per_cell_rates	shape [20, 53, 5]→[20, 47, 5]; the in-block rule prose "all 53 gated cells"→"all 47" (item-2 scope judgment: the count inside the same spec whose shape changes — flagged for the fidelity referee)
3	thresholds.family_c.gate_partition + report-only	C1 demoted: gate_eligible: [fingerprint.elimination_plus2pp], 1/1; report_reasons: fingerprint_reversal_not_realized; C1 ordering fields RETAINED under fingerprints.c1; disclosure note added
3	candidate_procedure.pass_rule	annotated: family C now GATES the C2 fingerprint only
4	family_c.check	"BOTH fingerprints reverse"→"the C2 fingerprint reverses (C1 demoted…)"; trailing sentence cured "both reverse"→"C2 reverses… C1's non-reversal is structural" (a §7-item-4 consequential cure — the old sentence became false under the ratified amendment; flagged for the fidelity referee)
5	gate_w1.covers	family A "47 gated / 58 report-only"; family C "1 gated / 1 report-only (C1 demoted by amendment 2)"; roll-up "48 gated / 84 report-only"
6	certification_scope.certifies	48-cell text + report-only pending bridges + "reverses the C2 fingerprint" (grammar cure: "orderings"→"ordering")
7	gate_w1.history	amendment-2 entry appended; §8 content verbatim; referee_round extended with fixes_2/re_verification keys to record the real five-round ceremony (template had three — flagged)

Zero threshold movement (retained values byte-equal to the frozen floor); thresholds.family_b, protocol statistic/pass_rule/estimator, governance, frame_pin, holdout_basis untouched.

Tests

• Partition bindings polarity-flipped: test_gate_w1_derivations.py (47/58 partition + two-reason ledger + retained tolerances; OC 0.9344/0.9623 recomputed from live tolerances + frozen sigmas, AND the 53-cell pre-amendment OC still recomputed via retained_tolerances — the frozen artifact is untouched; schema [20,47,5]; family-C 1/1 + check/pass_rule bindings; covers/certification 48/84).
• Proposal-era guards flipped: test_gate_w1_amendment2_proposal.py — the untouched-guard is now the §7 subset master-compare (no sibling gate moves; stable pre- and post-merge); partition/OC tests bind live gates.yaml to the ledger's "after" state, reconstructing the 53-cell "before" basis from retained_tolerances.
• Amendment-1-era bindings updated with attribution: the a1 OC-invariance test reconstructs its 53-cell basis via retained_tolerances; the a1 flip's family-A/C byte-identity-to-master claim is restated as the always-true invariant (no gated-or-retained VALUE moves; fingerprint anchors/orders unchanged) since a2 legitimately rescopes those families.
• --transition legacy-manifest regen: run on the clean committed tree, idempotent (second run = no diff).
• Full always-runnable suite: 1089 passed / 41 skipped / 0 failed (disposable in-worktree env); tier counts unchanged (renames only); test_tier_policy green under full collection; black + ruff clean.

Record notes

• The amendment squash subject on master retains "(draft — ceremony pending)" from an unrefreshed PR title; the ceremony completed before the merge (adversarial 4953763322 → fixes 4953911698 → verification 4954135376 → fixes-2 4954294003 → re-verification RATIFY AS-IS 4954346300). Standing correction adopted: refresh PR titles with bodies before ratification merges.
• Flip authored partly by the orchestrator directly (session-limit churn killed the flip lane repeatedly mid-edit); every edit is §7-derived and the fidelity referee reviews the result, not the authorship path.

Ceremony next step: flip-fidelity referee before merge. Do not merge until that round posts MERGE.

🤖 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 12:52pm

[MaxGhenis]

Verdict: MERGE.

Flip-fidelity round on w1-amendment2-flip head d90fcac (single commit on c50f8bb), under the amendment-1 flip-fidelity bar (#165 comment 4952245964: §7 verbatim yaml-paths 1:1, zero threshold movement, retained values byte-equal, mutation-fragile, --transition idempotent). Everything below I recomputed myself in a detached worktree against the ratified source of truth (c50f8bb:docs/amendments/gate_w1_amendment_2_family_a_concept_cells.md §7+§8), the frozen floor runs/gate_w1_floors_v1.json, a disposable in-worktree env, and the live GitHub ceremony record. The flip applies the ratified §7 exactly, moves zero thresholds to machine identity, the §8 history entry is token-verbatim, the surface is genuinely mutation-fragile, --transition is idempotent, and CI is fresh-green. No blocker; two non-blocking coverage notes at the end.

---

1. §7 → yaml path census (my own git diff c50f8bb..d90fcac -- gates.yaml)

15 textual hunks, 33 changed leaf paths on a full-doc deep compare — every one inside gates.gate_w1; ZERO outside; ZERO unmapped. Sibling gates gate_1 / gate_2 / gate_3 / gate_m4 and every non-gates top-level key are deep-equal to c50f8bb.

diff hunk (@@ -line)	changed paths	§7 item
-3707	covers roll-up 53/52→47/58, 2 gated→1/1, 55/77→48/84	5
-3891	family_a.views.earnings_participation.tolerances −2 (18-24 f/m)	1
-3907	…earnings_participation.derivations.rules −2	1
-4017	…marital_share.tolerances −2 (married.65+ f/m)	1
-4041	…marital_share.derivations.rules −2	1
-4085	…coresident_spouse.tolerances −2 (65+ f/m)	1
-4105	…coresident_spouse.derivations.rules −2	1
-4132	family_a.report_only +6, +report_reasons, +retained_tolerances, +demoted_disclosure_note	1
-4268	family_a.faithful_candidate_oc n 53→47, p_seed 0.922→0.9344, p_gate 0.9481→0.9623, basis_note	1
-4522	family_c.check BOTH→C2-only (+ trailing cure)	4
-4557	family_c.candidate_procedure.pass_rule annotated C2-only	3
-4608	family_c.gate_partition 2/0→1/1, +report_only, +report_reasons, +demoted_disclosure_note	3
-4684	protocol…per_draw_per_cell_rates.shape [20,53,5]→[20,47,5] + .rule prose	2
-4757	certification_scope.certifies 55→48, C2-only, +report-only-pending	6
-4953	gate_w1.history append a2 entry	7 / §8

The two demoted_disclosure_note additions map to §7-1 ("retained… under the report-only disclosure") and §7-3 ("retained under report-only"); the note schema is the ratified amendment-1 precedent (family_b.demoted_disclosure_note already exists at c50f8bb). No hunk is unmapped.

2. Zero threshold movement (machine-exact, recomputed independently)

• Six retained tolerances byte-equal to the pre-flip gated values and to the frozen floor's derived round(mean + 4·sd, 3): 0.211 / 0.221 (18-24 f/m), 0.163 / 0.084 (married.65+ f/m), 0.168 / 0.094 (coresident.65+ f/m). All 53 (47 live + 6 retained) reproduce the floor derivation exactly.
• OC recompute on the draw-noise-free half-normal basis (2·Φ(tol/σ)−1, p⁵+5p⁴(1−p)), σ from the frozen floor: 47-cell → p_seed 0.9344 / p_gate 0.9623 (= contract); 53-cell reconstructed via retained_tolerances → 0.922 / 0.9481 (= the FROZEN floor's own untouched OC). The floor JSON is byte-untouched and still lists all 53 as gate_eligible — the amendment rescopes the CONTRACT, not the floor, exactly as the disclosure note claims.
• C1 untouched: fingerprints.c1 deep-equal (anchor_values, required_representative_order, swap_pair, mechanism, psid_frame_order all identical); c2 identical.
• No sibling movement: family_b, protocol.{statistic,pass_rule,estimator,floor bindings} (only per_draw_per_cell_rates changed), governance, frame_pin, holdout_basis all deep-equal. Roll-up checks out: 47 A + 1 C = 48 gated; 58 A + 25 B + 1 C = 84 report-only.

3. The six flagged judgment calls

• (a) item-2 rule prose "all 53"→"all 47" — RATIFIED-AUTHORITY-CONSISTENT. The prose is a child of the exact block §7-2 mutates (per_draw_per_cell_rates); it restates the very count the shape encodes. Leaving it 53 would make the block internally false. Faithful propagation within the named path, not new scope.
• (b) item-4 trailing-sentence cure ("If transport is real C2 reverses… C1's non-reversal is structural") — RATIFIED-AUTHORITY-CONSISTENT. Same field (family_c.check) §7-4 rewrites; the old "both reverse; if cosmetic they do not" became false under the amendment. The cure matches §9's own "C1's non-reversal is structural" framing.
• (c) item-6 grammar cure "orderings"→"ordering" — RATIFIED-AUTHORITY-CONSISTENT. Forced number agreement: §7-6 singularizes "reverses BOTH fingerprints"→"the C2 fingerprint", so "anchor orderings"→"anchor ordering" is mechanical.
• (d) §8 referee_round extended with fixes_2/re_verification — RATIFIED-AUTHORITY-CONSISTENT. The §8 template's three keys carry # filled at ceremony placeholders; the real ceremony ran five rounds. I verified all five live: 4953763322 AMEND → 8e713ef/4953911698 fixes → 4954135376 FURTHER-FIXES → dd913ac/4954294003 fixes-2 → 4954346300 RATIFY AS-IS. Cramming five into three would lose ceremony record; extending faithfully records it (a1 precedent).
• (e) a1-era test updates (OC basis reconstruction via retained_tolerances; byte-identity restated as value-level zero-movement invariant) — RATIFIED-AUTHORITY-CONSISTENT. a2 legitimately changes family A/C, so a1's "byte-identical to master" is no longer the right invariant; the durable invariant both amendments guarantee is "no gated-or-retained VALUE moves + fingerprint anchors/orders unchanged." The restatement does not weaken the guard — mutations M2 (drop retained tol) and M8 (move retained value) are both caught by exactly these reconstructed-basis assertions. Reconstruction from retained_tolerances is valid precisely because I verified those values byte-equal.
• (f) orchestrator-authored flip (lane churn) — RATIFIED-AUTHORITY-CONSISTENT (no drift). Authorship is irrelevant and I found none of its risks realized: every hunk maps to §7 or a consistent cure, every number recomputes, §8 is token-verbatim, zero value movement, no editorializing, no scope creep, no unmapped hunk. The edits are exactly what a mechanical §7 application yields.

All six: RATIFIED-AUTHORITY-CONSISTENT. None is an overreach.

4. History entry (§8)

gate_w1.history gains exactly one entry 2026-07-12-w1-family-a-concept-cells (lock + a1-di-bands entries unchanged). Token-level compare of the applied content vs the ratified §8 draft: identical except one mid-token line-wrap — the doc split marital_share.married.65+| / female across a line boundary (doc line 644); the flip correctly rejoined it. All six demoted cell names verified intact, zero pipe-space-sex corruption. (The brief anticipated two wrap repairs; the other candidate wraps — ~0.577 F /, 18-24⏎participation, 0.922 / — all fold to correct text needing no repair, so I count one genuine token repair.) Ceremony fields all correct and live-verified: shas 8e713ef/dd913ac exist with the right subjects; ratified c50f8bb = PR #169's merge commit; RATIFY-AS-IS 4954346300; the stale-squash-title note is present and accurate (PR #169's title does carry "(draft — ceremony pending)").

5. Test polarity & mutations

• Always-runnable suite (-m "unit or artifact", disposable env): 1089 passed / 41 skipped / 0 failed — matches the target exactly. (The unmarked pytest tests collects 2603 and pulls in integration_psid/oracle_policyengine against local data; those are excluded from always-runnable, as on CI.)
• a2-proposal subset master-compare (test_gates_yaml_flipped_per_section7_and_moves_no_sibling): passes now against origin/master=c50f8bb; its set(live)==set(master) + per-sibling equality is invariant pre- and post-merge (gate_w1 excluded from the loop; siblings never move). Also green: test_gate_w1_flip_leaves_locked_siblings_byte_identical, test_flip_demotes_all_ten_family_b_cells_vs_master.
• Mutation battery — 8/8 CAUGHT (each applied to live gates.yaml, tree restored after):

#	mutation	result	representative failing test
M1	re-add earnings_participation.18-24|female to a gated view	CAUGHT (6)	test_gate_w1_family_a_power_cap_and_partition_47_58
M2	drop retained tolerance married.65+|male	CAUGHT (5)	test_family_a_oc_recomputes_before_and_after…
M3	p_gate_pass_4_of_5 0.9623→0.9624	CAUGHT (2)	test_gate_w1_family_a_oc_recomputes_from_tolerances_and_sigmas
M4	restore C1 to gate_eligible (n=2)	CAUGHT (3)	test_gate_w1_family_c_orderings_derive_from_committed_anchors
M5	covers roll-up 48→49 gated	CAUGHT (1)	test_gate_w1_covers_and_certification_scope
M6	history entry ratified sha c50f8bb→c50f8bc	CAUGHT (21)	test_legacy_manifest_matches_working_tree + test_contract_identity…
M7	revert cube shape [20,47,5]→[20,53,5]	CAUGHT (1)	test_gate_w1_protocol_is_the_ratified_k20_estimator_stream_9100
M8	move retained value coresident.65+|female 0.168→0.167	CAUGHT (6)	test_family_a_oc_recomputes_before_and_after…

6. --transition + CI

• scripts/build_legacy_manifest.py --transition on the clean committed tree: no diff; rerun idempotent (no diff). Manifest contract_revision = c43b2be = the flip commit's gates.yaml blob sha; recorded sha256/n_bytes match the working blob (305709 bytes).
• Fresh gh pr view 174: mergeable: MERGEABLE, mergeStateStatus: CLEAN. Fresh gh pr checks 174: exit 0 — pytest 3.11 + 3.13 pass, build + lint pass, Vercel pass, deploy skipping.

Non-blocking coverage notes (not AMENDs — the flip is faithful; these are for a follow-up)

1. M6 is caught by the byte guards, not a semantic binding. Any gates.yaml tamper trips test_contract_identity + test_legacy_manifest_matches_working_tree (blob/sha256 vs HEAD/manifest). No live test asserts the a2 history entry's content (ratified sha, ceremony comment ids, cell names) the way test_gate_w1_amendment1_flip_history_and_ceremony_record binds a1's. The a2 content is verbatim-correct (§4 above, verified independently), and the byte guards satisfy the brief's "must fail a test." A dedicated …amendment2_flip_history_and_ceremony_record test would close the asymmetry.
2. Two accepted string-binding limits already noted at ratification (re-verification 4954346300) live in the doc, not the contract, and don't affect the flip.

Recomputed numbers, path census, mutations, ceremony record, --transition, and fresh CI all hold. MERGE. (Merge is the orchestrator's act — no ready/merge from me.)