gate_w1 amendment 3 flip: demote hh_size quad + pair-scope C2 (ratified cb1ef63)

[MaxGhenis]

Applies the RATIFIED gate_w1 amendment 3 — 2026-07-13-w1-hh-size-and-c2-pair, ratified by merge of PR #180 (squash cb1ef63) after the five-round ceremony — to gates.yaml, per the string-bound §7 flip text of docs/amendments/gate_w1_amendment_3_hh_size_and_c2_pair.md.

§7 → yaml application table

§7 item	Path	Applied
1	thresholds.family_a views → report_only + report_reasons + retained_tolerances	hh_size quad moved; uniform reason no_permitted_entry_state_lever_reaches_cell; tolerances retained verbatim (0.191/0.191/0.174/0.184); demoted_disclosure_note_amendment_3 with per-cell channel attribution + per-cell truth (size-2 stays gated; size-4 cleared c3 seed 1)
1	family_a.faithful_candidate_oc	47→43 cells, 0.9344→0.9403, 0.9623→0.9684
2	protocol...per_draw_per_cell_rates	shape [20,47,5]→[20,43,5]; in-block rule prose "all 47"→"all 43" (the a2 item-2 scope judgment, same class — flagged)
3	family_c.fingerprints.c2	gated_check_amendment_3 (required_swap_realised on the pair only) + report_only_legs_amendment_3 (cap_150k/+1pp; machine reason; pinned-frame disclosure carrying "ON THE PINNED FRAME, NOT a universal claim" + f<0.0476); orderings retained verbatim; gate_partition unchanged 1/1
3	candidate_procedure.pass_rule	amendment-3 annotation appended (pair gates; full ordering still computed and published)
4	family_c.check	pair-scoped text per §7
5	gate_w1.covers	family A 43/62; family C 1/1 (pair-scoped); roll-up 44 gated / 88 report-only
6	certification_scope.certifies	44-cell text; hh_size + cap_150k/+1pp legs report-only pending; "reverses the C2 elim<->+2pp swap"
7	gate_w1.history	amendment-3 entry with the real five-round record (fixes_2/re_verification keys per the a2 template extension) + the title-refreshed-pre-merge note

Zero threshold movement; family_b, C1 disposition, protocol statistic/pass_rule/estimator, governance, frame_pin untouched. Judgment calls (all same-class as a2 flip precedents, for the fidelity referee): the rule-prose count inside the schema block; the pass_rule annotation appended rather than replacing the a2 annotation (history-preserving); the disclosure notes as separate _amendment_3 keys (a2 notes byte-intact).

Tests

Partition bindings polarity-flipped (43/62 partition, three-reason ledger, OC 0.9403/0.9684 + reconstructed 47-cell 0.9344/0.9623 and 53-cell 0.922/0.9481 via retained_tolerances, schema [20,43,5], family-C pair-scope bindings incl. the legs + disclosure, covers/certification 44/88); a2-era live-state bindings reconstruct their bases via retained_tolerances with attribution; the a3 proposal's untouched-guard became the §7 subset master-compare (stable pre/post-merge); the no-self-rescue test reconstructs the scoring-time 47-surface. --transition idempotent on the committed tree. Full always-runnable suite 1280/41/0; tier counts unchanged (2,797 total); black+ruff clean.

Flip authored inline by the orchestrator (session churn killed the flip lane repeatedly); every edit is §7-derived and the fidelity referee reviews the result.

Do not merge until the flip-fidelity 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 14, 2026 12:13am

[MaxGhenis]

Verdict: AMEND (1).

Flip-fidelity round on w1-amendment3-flip head f557abb (single commit on cb1ef63; merge-base with master = cb1ef63; master has since moved to 91f12f2 = the #181 M6 §2.8 docs merge — docs-only, zero file overlap with this flip, tests/tier_counts.json untouched on both sides, PR MERGEABLE/CLEAN against the moved base), under the amendment-2 flip-fidelity bar (#174 comment 4958900679). Everything below I recomputed myself in a detached worktree + disposable env against the ratified source of truth (cb1ef63:docs/amendments/gate_w1_amendment_3_hh_size_and_c2_pair.md §7+§8 — byte-identical at HEAD), the frozen floor runs/gate_w1_floors_v1.json, and the live GitHub ceremony record. The flip applies §7 exactly (zero unmapped hunks), moves zero thresholds to machine identity, the §8 entry is token-verbatim, the five-round ceremony record live-verifies in full, the surface is mutation-fragile 8/8, --transition is idempotent, and CI is fresh-green. One AMEND item — an in-block prose propagation this flip's own cited precedent requires — then this is merge-grade.

AMEND 1 — family_a.faithful_candidate_oc.basis_note left stale on the 47-cell surface

The flip updates the block's three numeric fields per §7-1 (47→43, 0.9344→0.9403, 0.9623→0.9684 — all verified) but leaves basis_note byte-identical, so the block now reads n_gated_cells: 43 next to three false self-descriptions: "seed pass = product over the 47 gated cells", "RECOMPUTED on the 47-cell gated surface after amendment 2 …", "Recomputed from the 47 retained view tolerances …". This is exactly the class the a2 flip cured in this very field (d250d4b rewrote "53 gated cells"→"47 gated cells" and recorded the a2 recompute; the a2 fidelity round passed that as part of the §7-1 hunk), and the class this flip itself cured one block over (the flagged "all 47"→"all 43" schema-rule prose, PR-body judgment call (a), whose a2-precedent rationale — leaving it stale "would make the block internally false" — applies verbatim here). No test binds the note; no number moves; the ceremony never discussed basis_note (swept all five comments + the ratified doc), so no ratified text blocks the propagation — §7-1's "the faithful_candidate_oc block updates" plus the a2-flip treatment define it. Fix: propagate the three counts to 43 and record the a3 recompute in the note (mirror the a2 sentence pattern; keep the frozen-floor 53-cell sentence), rerun --transition (contract_revision tracks the new blob), suite stays green. A delta re-check of the single edited scalar then suffices.

1. §7 → yaml path census (my own git diff cb1ef63..HEAD -- gates.yaml)

13 textual hunks, 30 changed leaf paths on a full-doc deep compare — every one inside gates.gate_w1; ZERO outside; ZERO unmapped. Siblings gate_1/gate_2/gate_3/gate_m4/gate_m6 deep-equal AND raw-text byte-identical (gate_m6 byte-identical as required); every non-gates top-level key deep-equal.

hunk (@@ -line)	changed paths	§7
-3707	covers: A 47/58→43/62; C parenthetical (pair re-scope, legs report-only within C2); roll-up 48/84→44/88	5
-4047	views.hh_size_share.tolerances −4 (keeps .2: 0.123); derivations.rules −4 (keeps .2)	1
-4118	family_a.report_only +4 (the quad)	1
-4126	report_reasons +4 → no_permitted_entry_state_lever_reaches_cell (uniform)	1
-4139	retained_tolerances +4 (0.191/0.191/0.174/0.184; a2 view/k/rounding schema)	1
-4160	+demoted_disclosure_note_amendment_3 (a2 note byte-intact above it)	1 + call (c)
-4295	faithful_candidate_oc 47→43, 0.9344→0.9403, 0.9623→0.9684 (basis_note NOT propagated → AMEND 1)	1
-4555	family_c.check → §7-4 replacement verbatim + in-path cures (below)	4
-4597	candidate_procedure.pass_rule += a3 annotation (a2 text a verified strict prefix)	3 + call (b)
-4648	c2 +gated_check_amendment_3, +report_only_legs_amendment_3	3 + call (c)
-4744	per_draw_per_cell_rates.shape [20,47,5]→**[20,43,5]** + rule "all 47"→"all 43"	2 + call (a)
-4817	certification_scope.certifies → §7-6 replacement verbatim + narrowed "reverses" clause	6
-5127	history += the a3 entry	7/§8

The §7-4 and §7-6 quoted replacement texts verify verbatim inside the applied fields (modulo the file's ascii <->/-- convention); the a3 disclosure note carries every ratified figure (0.111 roster ceiling, 15-entry ceiling, 0.54/0.46, 0.75, the 0.98 mirror, 0.67, 0.78, size-2 stays gated at 0.123 clearing c3 5/5, size-4 cleared c3 seed 1, sizes 1/3/5plus 5/5-fail ×3) plus the explicit anti-universal ("not a claim that the cells are unreachable by any future model class").

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

• Four hh_size retained tolerances: byte-equal to the pre-flip gated values AND the frozen floor's round(mean + 4·sd, 3): 0.191 / 0.191 / 0.174 / 0.184. hh_size_share.2 stays gated at 0.123 = its floor derivation. Six a2-retained unchanged and floor-exact (0.211/0.221/0.163/0.084/0.168/0.094). All 43 live tolerances reproduce the floor derivation. Floor file untouched (last commit W1 transport gate: floors evidence base #154), still lists all 53 gate-eligible and keeps its own 53-cell OC.
• OC recompute (half-normal basis, floor σ): 43-cell → 0.9403 / 0.9684 (= contract, exact); 47-cell reconstructed via retained_tolerances → 0.9344 / 0.9623 (= the a2 contract); 53-cell → 0.922 / 0.9481 (= the frozen floor's own untouched block). Spot per-cell pass-probs bit-match the floor.
• C2 orderings/anchors untouched: anchor_values (21/18/5/1 — the provision deltas), required_representative_order, psid_frame_order, swap_pair, mechanism, both τ fields, both sources all machine-EQUAL; only the two _amendment_3 annotation keys are added. gate_partition unchanged 1/1. c1 deep-equal; family_b deep-equal; protocol changed only in per_draw_per_cell_rates.{shape,rule} — statistic/pass_rule/estimator/undefined_draw_rule/floor bindings equal (the protocol's 53-cell floor-mirror OC correctly untouched); governance/frame_pin/holdout_basis equal; both a2 disclosure notes byte-intact; roll-up recomputes 44 gated / 88 report-only from the structures (43+1 / 62+25+1).

3. The three flagged judgment calls (+ two in-path cures)

• (a) schema rule-prose "all 47"→"all 43" — RATIFIED-AUTHORITY-CONSISTENT. Child prose of the exact block §7-2 mutates, restating the count the shape encodes; the a2 item-2 precedent verbatim. (And see AMEND 1: the identical principle left unapplied in the OC block.)
• (b) pass_rule annotation APPENDED — RATIFIED-AUTHORITY-CONSISTENT. The a2 text is a verified strict prefix (append-only, byte-intact), continuing the lock→a2→a3 layered-annotation style the a2 flip established; the appended text states the operative rule unambiguously ("PASS iff required_swap_realised on [payroll_plus_2pp, elimination] … The full-ordering statistic is still computed and published; only the pair gates") — §7-3's re-scope operationalized at the pass-definition site.
• (c) separate _amendment_3 disclosure keys — RATIFIED-AUTHORITY-CONSISTENT. The a1/a2 note-schema precedent; both a2 notes byte-intact (editing them would corrupt a ratified record); §1/§5's required per-cell channel-attribution + per-cell pass records need a contract home, and this is the precedent locus.
• Two in-path cures inside family_c.check (not flagged; the a2 item-4 cure class): the C1 parenthetical became a standalone sentence (content-preserving — fingerprint_reversal_not_realized, amendment 2, ordering + PPI-NRA gap publish all retained; the "structural" substance lives byte-intact in the a2 demoted_disclosure_note), and "and no OTHER adjacency changes" was dropped (keeping it would contradict the ratified re-scope — the certified correction mechanically moves the cap adjacency, Q10). Both forced by the §7-4 rewrite within its named path. Cosmetic note: gated_check_amendment_3 says "provision_deltas above" in §7-3's own vocabulary; the yaml key carrying those deltas is anchor_values (machine-EQUAL).

4. §8 history entry + ceremony record (all live-verified)

• Content block token-verbatim vs the ratified §8: identical modulo exactly three mid-token line-wrap rejoins — (0.191/0.191/0.174/0.184), PER-CELL, report-only, — the a2 bar's wrap-repair class (a2 had one; a3 has three), zero token added/removed/altered otherwise. id/proposed/flipped_live == the draft.
• referee_round extended review/fixes/verification + fixes_2/re_verification = the a2 template-extension precedent exactly (the a2 entry carries the same five keys, and a2 likewise expanded review from the short template to the real record).
• Five comments fetched and matched: 4962052314 AMEND THE AMENDMENT (the merge-base proof that amendment3_risks = exactly the acted-on set and pre-dates c3; F1 = the a2 defect class re-instantiated, refuted by order_at_f_deployed_0p22; F2 = fertility terminology) → fixes 60f6ee5 / 4962297637 (ledger cap_150k_impossibility with pinned_certified_frame_only + is_universal_claim: false; BANNED_UNIVERSAL_C2_PHRASINGS — the constant exists in the module, 11 phrasings; floor→ceiling; 26 bindings, 16 mutations) → 4962578395 FURTHER FIXES ×2 (V1 the PR-body/squash universal outside the battery's reach; V2 the §2a clause; M4/M4b blindness; machine reason ACCEPTABLE-WITH-DEFINITION, final) → fixes-2 e3c63de + master-merge 1d51b77 / 4962844009 (pinned_A_over_B_needs_f_below_round4 = 0.0476 bound; "(<0.5)" gloss + conjunct restored; guard extended, catches demonstrated) → 4963441188 RATIFY AS-IS (0.0476 re-derived and adjudicated; the fresh-synonym honest limit disclosed; merge anatomy clean). All three shas exist with matching subjects; cb1ef63 = PR gate_w1 amendment 3: demote the hh_size quad + pair-scope C2 (ratified after five-round ceremony) #180's squash merge (single parent 95a3e05); the squash subject carries "(ratified after five-round ceremony)" with no draft residue — the title-refresh note is accurate.

5. The pinned-frame disclosure + universal sweep

report_only_legs_amendment_3: legs: [cap_150k, payroll_plus_1pp], the ratified machine reason, and a disclosure carrying "ON THE PINNED FRAME, NOT a universal claim", "f < 0.0476" (recomputed: 0.0806/1.694 = 0.047580… → 0.0476, the ceremony's round-4 value), the order_at_f_deployed_0p22 caveat (A/B≈0.22, f≈0.22 realises the swap AND holds the full order), rank-2/0.373/16.7yr/1.694 (all = the ratified §8/ledger figures), restoration-enumerated-empty, and the publish-with-every-fresh-run clause — and the live suite string-binds both load-bearing phrases (tests/test_gate_w1_derivations.py family-C test asserts "ON THE PINNED FRAME, NOT a universal claim" and "f < 0.0476"). Swept every added line of the gates.yaml diff: zero of the 11 BANNED_UNIVERSAL_C2_PHRASINGS present; every "no frame"/"cannot hold"/"internally inconsistent" instance is the scoped or anti-universal form ("not one no frame can", "ON THE PINNED FRAME", "the pinned transport target cannot hold"). No universal reintroduced anywhere.

6. Tests + mutations

• Always-runnable suite (-m "unit or artifact", disposable env): 1280 passed / 41 skipped / 0 failed — the target exactly; 2797 collected = the ratified tier total, tests/tier_counts.json untouched.
• Reconstructed-basis tests run verbosely and pass, and they genuinely recompute (per-cell half-normal products from floor σ, asserted to 0.9403/0.9684 live, 0.9344/0.9623 via the a3-retained quad, 0.922/0.9481 via all ten retained; the no-self-rescue test reconstructs the scoring-time 47-surface as live 43 + quad and re-checks the c3 per-seed records). The a2-era modules reconstructing their bases via retained_tolerances is the a2 bar's judgment-call-(e) class — valid precisely because I verified those values byte-equal (§2).
• Mutation battery — 8/8 CAUGHT (each applied to live gates.yaml, tree restored byte-exact after):

#	mutation	result	failures (semantic)	representative failing test
M1	re-gate hh_size_share.1 in view tolerances	CAUGHT	8 (7)	…family_a_power_cap_and_partition_43_62
M2	drop retained hh_size_share.4	CAUGHT	7 (6)	…oc_recomputes_before_and_after_on_frozen_floor
M3	p_gate_pass_4_of_5 0.9684→0.9685	CAUGHT	5 (4)	…family_a_oc_recomputes_from_tolerances_and_sigmas
M4	restore the 4-element C2 check (revert check + delete both _amendment_3 keys)	CAUGHT	3 (2)	…family_c_orderings_derive_from_committed_anchors
M5	covers roll-up 44→45	CAUGHT	2 (1)	…covers_and_certification_scope
M6	history ratified sha cb1ef63→cb1ef64	CAUGHT	1 (0 — byte guard)	test_legacy_manifest_matches_working_tree
M7	cube shape →[20,47,5] (+rule)	CAUGHT	3 (2)	…protocol_is_the_ratified_k20_estimator_stream_9100
M8	retained hh_size_share.1 0.191→0.190	CAUGHT	6 (5)	…family_a_power_cap_and_partition_43_62

7. --transition + CI

• scripts/build_legacy_manifest.py --transition on the clean committed tree: no diff; rerun idempotent (no diff). Manifest contract_revision = d68449f = the flip commit's gates.yaml blob sha; recorded sha256/n_bytes match the working blob (349953 bytes).
• CI: the initial Pages run on f557abb was CANCELLED (not failed) by the repo-global concurrency: group: "pages" when master's M6 design amendment 3: the scored-run harness (§2.8, unblocks the gate_m6 run) #181 run started 3 s later; I re-ran it → build pass, lint pass, deploy skipping (the PR-event pattern, as on gate_w1 amendment 2 flip: demote family-A concept cells + C1 to report-only (ratified c50f8bb) #174); pytest 3.11 + 3.13 passed first time. Fresh: gh pr checks 182 exit 0; mergeable: MERGEABLE, mergeStateStatus: CLEAN, still DRAFT.

Non-blocking notes

1. The history-content binding asymmetry persists (the a2 round's note 1, still open): M6 is caught by the manifest byte guards only; a1 has a dedicated flip-history content test, a2/a3 don't. Today the §8-doc string bindings + my token-compare + the byte guards cover it transitively; a dedicated …amendment3_flip_history_and_ceremony_record test (and a2's) would close it.
2. governance.amendment_rules.description_claims_exactly_the_scored_surface still enumerates the lock-era "53 family-A + 10 family-B + 2 family-C" surface. §7 lists governance as untouched and the a1/a2 flips left it identically, so this flip is right not to touch it — but a future hygiene amendment could re-point the enumeration at covers so the rule stops carrying a stale instantiation.

AMEND (1): propagate basis_note to the 43-cell surface (the a2-precedent treatment), rerun --transition, and re-request this round — a delta re-check of the single edited scalar suffices. Everything else recomputes, live-verifies, and is merge-grade. (Merge is the orchestrator's act — no ready/merge from me.)

[MaxGhenis]

Verdict update: MERGE. AMEND 1 resolved in 84565d4; delta re-check clean, all recomputed in a fresh worktree + env.

• Scope — exactly the fix, nothing else. Deep YAML compare f557abb..84565d4 of gates.yaml = one changed leaf: .gates.gate_w1.thresholds.family_a.faithful_candidate_oc.basis_note. The only other file is runs/legacy_manifest_v1.json (the --transition fields). No OC scalar, no sibling, no other gate_w1 path moved.
• Prose claims-exact. All three stale current-surface self-descriptions propagated 47→43 ("product over the 43 gated cells", "RECOMPUTED on the 43-cell gated surface after amendment 3 demoted the 4 hh_size cells", "from the 43 retained view tolerances"). The only remaining 47 is the correct a2 historical record ("following the amendment-2 recompute to 47 cells / 0.9344 / 0.9623"), recorded following the a2 sentence exactly as intended. The FROZEN-floor 53-cell sentence is byte-intact ("keeps its own 53-cell faithful_candidate_oc 0.922 / 0.9481"); "the amendments re-scope the CONTRACT, not the floor" correctly pluralized; test-binding pointer intact. OC scalars unchanged at the ratified 43 / 0.9403 / 0.9684.
• --transition idempotent (two runs, 0 diff, clean tree); manifest contract_revision = 1efbf09 = HEAD:gates.yaml blob; sha256/n_bytes blob-exact (349996 bytes).
• Suite in the disposable env: always-runnable 1280 passed / 41 skipped / 0 failed (unchanged); test_gate_w1_derivations.py + test_legacy_manifest.py 57/57.
• Fresh CI on 84565d4: gh pr checks 182 exit 0 (build/lint/pytest 3.11+3.13/Vercel pass, deploy skipping — clean PR-event run this time, no concurrency cancellation); mergeable: MERGEABLE, mergeStateStatus: CLEAN.

The §7 1:1 application, zero-threshold-movement, the three judgment calls, the §8 token-verbatim entry + five-round ceremony record, the pinned-frame disclosure, and the 8/8 mutation battery all held in the prior round and are untouched by this prose-only delta. Merge-grade. (Merge is the orchestrator's act.)