W1 candidate 1: first full transport deployment (registration 4950931131) by MaxGhenis · Pull Request #162 · PolicyEngine/microcosm-dynamics

MaxGhenis · 2026-07-12T12:35:56Z

One-shot scored run of W1 candidate 1 — the first full transport deployment against the locked gate_w1 contract. Registered at issue #42 comment 4950931131. Publishes regardless of the verdict; no holdout tuning. The certified PSID-estimated generators (gate-1 earnings, gate-2a/2b marital CANDIDATE_16, gate-2b household CANDIDATE_9, gate_m4 disability, #115/#117 reform ledgers) are deployed on the certified populace CPS frame (us-4.18.8, sha c2065b64…, sha-verified) — a frame none was fitted to.

Verdict: FAIL (0 / 5 seeds)

gate_pass = false — family A 0/5 seeds, family B 0/10 cells, family C neither fingerprint reverses. This lands inside the registration's forecast P(gate ≥ 4/5) = 0.05–0.20. The value is the per-family decomposition and the fingerprint diagnostic.

Per-seed verdict table (family A, 53 gated cells)

seed	cells pass	cells fail	seed pass
0	28/53	25	❌
1	28/53	25	❌
2	32/53	21	❌
3	26/53	27	❌
4	24/53	29	❌

Conformance: identity_candidate = false, max across-draw sd 0.475 — the candidate regenerates every scored column per the regenerated_surface rule (not the prohibited identity map, which would score 0 with zero dispersion).

The fingerprint reversal results (family C — the highest-value read)

Both fingerprints were re-run on the transported frame changing only the frame (the committed #115/#117 ledgers run verbatim; the earnings-source readers are swapped to gate-1-transported careers). Neither reverses to the anchor ordering.

fingerprint	PSID "before"	required "after"	deployed	reversed?
C1 PPI↔NRA	`PI > NRA > PPI > COLA`	`PI > PPI > NRA > COLA`	`PI > NRA > PPI > COLA`	NO (τ vs required 0.667, unchanged from PSID)
C2 elim↔+2pp	`+2pp > elim > +1pp > cap`	`elim > +2pp > +1pp > cap`	`elim > cap-$150k > +2pp > +1pp`	NO (τ vs required 0.333)

C1 does not move at all — PPI stays below NRA→70; the transported AIME above the second PIA bend does not lift progressive price indexing past the NRA cut.
C2 shows the target swap but not a clean reversal — the required elimination > +2pp swap is realised (required_swap_realised = true): the heavier transported earnings tail lifts taxable-max elimination above +2pp, exactly the Roadmap: from incidence replication to a DYNASIM-class projection on one unified person-period panel #113 mechanism. But the same tail also lifts cap-$150k above both payroll rates, adding a second adjacency change, so the ordering is not the committed [elim, +2pp, +1pp, cap].
Robustness: the two orderings were identical across three independent transport constructions (level-anchored, rank-flat, and rank + transitory mobility). The transported frame carries 37.7% of taxable payroll above the wage base — an upper read (positive prime-age earners, no zero/low-earning years), well above the real ~17% and the PSID frame's 12.7%. A heavier tail favours reversal, so the non-reversal is conservative: a realistic tail would be no more likely to reverse.

Per the registration, "their failing to reverse would be the single most important diagnostic in the program." The diagnostic here: deploying the PSID-fit generators on the CPS frame does move the taxable-max ordering in the predicted direction (C2 swap realised) but does not cleanly reproduce either committed anchor ordering.

Per-family decomposition

Family A (CPS-observable joints, floor-priced) — modal failure by view (cell-fails summed over 5 seeds):

view	cell-fails	reading
`marital_share`	45	younger cohorts (25–34, 35–44) under-produce married: the synthetic-panel marital deploy starts everyone never-married and the hazards do not equilibrate over the exposure window (older 45–64 cohorts largely PASS)
`earnings_participation`	25	gate-1 has no sex covariate and fits ages 25–59; 18–24 / 62–69 extrapolate to the nearest fitted bin, overstating retirement-age participation
`hh_size_share`	25	the household generator composes ADULTS; committed `rate_a` is all-person (children live in larger households) + coresident-parent/multigen are under-generated on a cross-section
`coresident_spouse`	16	tracks the marital under-production at younger ages
`earnings_profile`	16	the 18–24 age-boundary extrapolation

19 earnings cells pass (including the p90/p50 and p50/p10 dispersion ratios), so the earnings regeneration is not degenerate. (The registration predicted earnings-profile joints as the modal class; the actual modal class is marital_share, reported as-is.)

Family B (SSA anchors, 10 gated) — 0/10. DI status is M4-simulated (never the frame's DI column or any SS proxy; conditioning columns {age, is_female} asserted populated). The M4 work-disability prevalence gradient is far flatter than the SSA disabled-worker beneficiary stock's steep concentration at older ages: e.g. di_prevalence.60-fra deployed 18.6pp vs anchor 45.4pp; conversion share ~6pp vs ~14pp. The gradient transports; the level and steepness do not — the second failure class the registration named.

Family C — both fingerprints fail to reverse (above).

Spec resolutions (documented; the locked text does not pin them)

The certified generators are PSID-panel simulators; deploying them on a CPS cross-section required synthetic-panel adapters. Eight resolutions are recorded in the artifact's spec_resolutions and the module docstring, including: generators fit once on the full PSID sample and frozen (nothing refits per W1 seed); earnings regenerated from the certified gate-1 cell marginal at (age-bin, terminal period 2022), not anchored on the frame's own earnings (which would be the identity); marital/household via synthetic panels with empty initial roster states; family-A surface scored on the adult universe (only hh_size_share's committed all-person rate_a differs in scope, disclosed); DI via Bernoulli(m4_prevalence[band,sex]); family-C careers transported with the committed permanent-variance share (0.467). One label note: gates.yaml calls the household generator "gate-2c" at the hh_size_share row; household composition is the gate-2b-locked CANDIDATE_9 everywhere else in the file, used here.

Reproduction

# 1. data-bound frame export (policyengine.py .venv, sha-verifies c2065b64…)
PYTHONPATH=src <pe.py-venv>/python scripts/export_frame_persons.py
# 2. fit + score (.venv-gate); family A chunked by seed per the memory guard
POPULACE_DYNAMICS_FRAME_PICKLE=scratch/frame_persons.pkl \
  <gate-venv>/python scripts/run_gate_w1_candidate1.py fit
#   then: seed 0..4 (parallel) + family-b + family-c, then assemble

Runtime: generator fit ~2 min (gate-1 marginals + ft 19s + hc 82s + m4); family A ~5 min/seed (5 parallel processes); family B ~2s; family C ~60s. Frame 166,302 persons / 340.0M weighted, sha-verified against the pin.

Tests

+11 tests (9 unit pure-logic, 2 artifact committed-json bindings — including the holdout-split binding that recomputes every gate seed's committed household sha256 with no h5). tier_counts.json refreshed in a CI-equivalent env (.venv, openpyxl present) per the #158 rule: unit 199→208, artifact 783→785. Always-runnable tiers green: 953 passed, 40 skipped (unit or artifact); test_tier_policy green over the full 2451-test collection. black + ruff clean.

Do not merge on my account; verification-before-record.

🤖 Generated with Claude Code

vercel · 2026-07-12T12:36:02Z

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
social-security-model	Ready	Preview, Comment	Jul 12, 2026 12:38pm

…131) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* gate_w1 amendment 1 proposal: the family-B DI bands (forensics-1 Q4) DRAFT proposal (ceremony step 1 of proposal -> referee -> fixes -> verification -> ratify -> flip). gates.yaml is UNTOUCHED; no threshold moves. W1 forensics 1 Q4 (runs/gate_w1_forensics1_v1.json, registration 4951218279, grading 4951430002) established, against the currently locked contract blob cd6411d, that the 8 family-B DI age-composition prevalence bands are unclearable by any contract-consistent candidate: the gate scores an M4 work-disability point-prevalence against the SSA disabled-worker beneficiary STOCK (DI ASR Table 19) across a concept bridge it never defined. The concept delta dominates (share 0.595; +21.3pp duration-accumulation vs +2.6pp M4 hazard shape at 60-FRA), and all 8 bands miss by 2.9x-21.9x their tolerances. Because family B is a conjunction, gate_w1 as locked is unpassable (faithful-candidate gate OC structurally 0). The proposal (docs/amendments/gate_w1_amendment_1_family_b_di_bands.md) analyses two options and recommends option (a): demote the 8 DI bands to report-only with machine reason concept_bridge_undefined_di_stock, leaving family B gating the 2 disability-conversion M4-simulated margins. Option (b) (re-anchor to a transportable work-disability prevalence) is deferred: no defensible anchor is archived, and the SSA-native re-anchor is blocked by the unarchived Supplement 4.C2 insured denominator. OC consequence: family A is untouched, so its half-normal characteristic is byte-identical (p_seed 0.922 / p_gate 0.9481, the 8 DI cells absent from the family-A OC machinery); the overall gate OC moves from a structural 0 to the family-A characteristic 0.9481 on the residual 57-cell surface (65 -> 57 gated, 67 -> 75 report-only). No-self-rescue: candidate 1 (runs/gate_w1_candidate1_v1.json, PR #162) stands FAIL under both surfaces -- it fails family A (0/5 seeds), the 2 RETAINED conversion cells (dev 8.30/8.07 vs tol 1.89/1.91), and family C -- so the demotion rescues nothing; the amendment is prospective; the forensics evidence is train/frame-side. tests/test_gate_w1_amendment1_proposal.py (12 tests, always-runnable, artifact tier) binds every proposal figure to the committed artifacts; tier manifest + README refreshed CI-equivalent (artifact 834 -> 846, total 2528). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * gate_w1 amendment 1 fixes round: referee findings (comment 4951701300) Adversarial round (PR #164 comment 4951701300, AMEND THE AMENDMENT) required extending the family-B demotion to all 10 gated cells; the 2 conversion margins fail the same candidate-independence test on committed evidence. A (blocker): all 10 family-B cells demoted to report-only. Conversion cells take concept_bridge_undefined_di_stock (same undefined DI-stock numerator, read at the FRA boundary) + conversion_level_match_never_certified (m4_disability_v1.json conversion_validation: ratio 0.267/0.322, 'never a level match'; window floor >=12.21/12.59 = anchor 14.1/14.5 - tol 1.89/1.91). Family B gates nothing; OC returns to 0.9481 x I(family C) on the residual 55-cell surface (65->55 gated, 67->77 report-only). Sections 1/3/4/4a/5/7/8/9 + ledger updated; option (b) conversion-anchor deferral row added. B: bound the operative flip text -- the section-7 enumerated demotion list, the section-7 roll-ups (55/77), and the section-8 history figures are parsed and cross-checked against the ledger/artifacts; proven to FAIL under the referee's named mutations. De-vacuoused the prose machine-reason assertion (checked outside the ledger fence). C: 0.595 restated as the aggregate with per-band range 0.023-0.891; CPS-ASEC option-(b) rationale corrected to built-from-frame (not the conditioning rule). Tier manifest refreshed CI-equivalent (artifact 846 -> 853; total 2535). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

…by its own evidence (#166) Carry the paper through every merged wave since the gate-2b pass (#149). Add the gate-2c candidate ladder and its independently verified pass (#155/#157, verification 4950580430): two candidates — the shortest of the family ladders — candidate 1 failing 1/5 with the assortative diagonal held 35/35 on the first attempt and the gate bound entirely on the earnings-blind first-marriage cells the registration had named in advance, candidate 2's one marginal-preserving modifier (byte-carry 2,500 cells at 0.0) passing 4/5. With it, every gate-2 tranche is passed: all four demographic gates locked AND passed (memo 4950584695). Add the disability gate (M4): floors #153 (round-1 restoring the occupancy- stock teeth via mixed-k), lock #159 (the start-wave weight-definition fix), candidate 1 passing 5/5 (#161) — the program's first five-of-five — verified with a both-constructions gap adjudication (4951142927), the estimand always the PSID self-report, never an SSA DI level (seven concept deltas). Add the transport gate (W1): design #151, floors #154 with the owned #158 tier-manifest process violation, lock #160, candidate 1 failing 0/5 (#162, C2 swap realised / C1 non-reversal robust / earnings not degenerate), forensics 1 (#163, the entry-state lever and the Q4 gate-design finding), and amendment 1 in full ceremony (#164/#165) — the program amending its own structurally-unpassable gate, all ten family-B cells demoted on committed evidence with anchors retained verbatim and the failed candidate left failed. This is the fourth recurring referee class: a gate can itself be wrong, and the same ceremony catches that too, after the lock. Add the interim benefit seam (#156: gross $41.35B -> net $36.02B, offsets enumerated, 4/4 expectations, engine-vs-transport gap 6e-8), stating the PolicyEngine-executes / Axiom-cross-validates division; and the deployment infrastructure (contract identity, env sidecars, a generic K=20 evaluator reproducing the three committed gate verdicts; the household candidate-9 registry port). Recompute the tally per the paper's conventions: forty-two registered runs across six gates (13/16/9/2/1/1), thirty-seven failures against five passes, six gates locked with FIVE ratified amendments (two gate 1, two gate 2, one transport), ten forensics rounds, six external-anchor replications, and fifty forecasts registered with forty-nine graded — the one open registration the transport gate's second candidate. Frozen sections byte-identical. Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

#169) * gate_w1 amendment 2 proposal: the family-A concept cells (forensics-2) DRAFT proposal (ceremony pending). Demotes 7 structurally-unclearable cells to report-only: the C1 fingerprint (fingerprint_reversal_not_realized), the two 18-24 participation cells (population_concept_delta_head_spouse_universe, a 22.1pp head/spouse-vs-all-person gap), and the four 65+ marital/coresident cells (cohort_vintage_hazard_frame_mismatch, divorce over-accumulation -- a cohort-vintage hazard/frame mismatch; six permitted-lever construction-attempts each fail). Zero threshold movement; gates.yaml untouched. Family-A faithful OC recomputes 0.922/0.9481 -> 0.9344/0.9623 on the residual 47 cells (removing 6 family-A cells, unlike amendment 1's invariance); overall gate OC 0.9623 x I(C2) on the 48-cell surface (55 -> 48 gated). No-self-rescue is over-determined: both candidate 1 (PR #162) and candidate 2 (PR #167) still fail family A (0/5 seeds on the 47) AND family C (C2 tau 0.333) -- the rescue set is empty. Consistency bindings in tests/test_gate_w1_amendment2_proposal.py string- bind the section-7 flip text/roll-ups and section-8 history (amendment-1 fix-B standard, mutation-checked). Tier artifact 879 -> 897 (CI-equivalent recount). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * gate_w1 amendment 2 fixes: per-cell grounds + retained-surface candor (referee 4953763322) Adversarial referee round (PR #169 comment 4953763322) ruled AMEND THE AMENDMENT; demotion set of 7 survives, grounds and candor corrected. Zero threshold movement; gates.yaml untouched; OC numbers unchanged (referee reproduced them exact). F1 (per-cell truth): the blanket "no candidate can clear any of the 7" was false -- candidate 1 PASSED coresident_spouse.65+|female 4/5 seeds. The 65+ quad now carries TWO machine reasons: cohort_vintage_hazard_frame_mismatch (married pair + coresident-65+|male, the stationary ceiling ~0.577 F / ~0.717 M the deployment cannot reach; married-65+|female caps 2-3/5 by ceiling-vs-window, Q1 entry-0 0.5767 vs Q6 entry-0.58 0.5728) and scored_duplicate_of_demoted_married_quantity (coresident-65+|female, deployed coresident/married ratio 1.0, straddles an easier window; demoted on duplicate/composition-blind grounds not impossibility). Per-seed pass records added to the doc + ledger + bindings; the c1 4/5 pass stated plainly. F2 (candor): the 48-cell surface is stated NOT KNOWN PASSABLE -- Q7 graded REFUTED for hh_size (all 5 cells 0/5 both candidates; "not closable by these two entry-state levers alone"); C2 is one bit-identical measurement (deployed cap_150k 16.7y vs Smith rank-4 1y; corrected-tail tau -1/3; robust only for the elim<->+2pp pair). Series forecast added: amendment-3 risks (5 hh_size + C2 cap_150k) pre-named, each reopenable only via the same forensics-then-ceremony standard. F3: section-3 sigma range corrected 3-8 -> 2.0-6.6 (0.11-0.44 ln vs 0.027-0.072 sigma; sub-3sigma end = the female cells that sometimes pass). F4: the referee's per-band-entry (0.4228/0.6004), entry-55 (0.4951/0.6090), and max-entry-bound construction-attempts incorporated into the 4c table (cite 4953763322). F9: section-4a Mermin sentence rewritten; C1 empirical legs qualified as one bit-identical measurement. Tests: 24 bindings (from 18), mutation-checked incl. the corrected sentences. Tier artifact 897 -> 903 (CI-equivalent recount). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * gate_w1 amendment 2 fixes 2: per-cell truth for §2c/§3.1 + MM4/MM6 bindings (verification 4954135376) The verification round (comment 4954135376, verdict FURTHER FIXES) ruled the scored_duplicate_of_demoted_married_quantity ground PRINCIPLED but caught two per-cell-false survivors introduced/inherited by the fixes commit, plus a stale body and one hh_size count (note N2). - §2c: 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, which candidate 1 clears 4/5 because the ~0.577 stationary ceiling sits INSIDE that cell's lower window on 4 of 5 seeds (lower edges [0.5478, 0.5759, 0.5717, 0.5769, 0.5813]). Rewritten to per-cell truth: ceiling-below-window holds for the married pair + coresident male; the coresident-female window straddles the ceiling, which is why its ground is duplication, not unreachability. Bound by test_section2c_no_blanket_65plus_impossibility_coresident_inside_window (MM4). - §3.1: dropped marital_share.never_married.25-34|male from the "5/5 for both candidates" list — candidate 1 clears it 1/5 (seed 1 0.1707 vs tol 0.192, flags [T,F,F,F,F]). 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; never_married is retained a fortiori. Bound by test_section3_1_5of5_fail_list_per_cell_true_excludes_never_married (MM6). - §4c: folded the verifier's principled ruling + deployment-contingent guard into the duplicate ground — 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 modeling coresidence separately, and re-gating then requires the standing forensics-then-ceremony path. Cites comment 4954135376. - N2: reconciled the amendment-3 hh_size risk count to FOUR (sizes 1/3/4/5+) in §1/§5/§8/§9 prose — Q7's joint lever clears only size-2. All five hh_size cells still fail 5/5. Bound in test_c2_record_bound_and_amendment3_forecast. Frontmatter + §11 checklist updated to record the verification round and this fixes-2 revision. Bindings 24 -> 26; artifact tier 903 -> 905, 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: prior FM1-8 / MM1-2 stay caught, and the MM4 / MM6 analogues the verification round found unbound are now caught. gates.yaml byte-identical to origin/master; the draft moves nothing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

…d C2 Proposal (DRAFT) to demote the four family-A household-size cells hh_size_share.{1,3,4,5plus} to report-only (uniform machine reason no_permitted_entry_state_lever_reaches_cell) and re-scope the family-C gated binary C2 from the full 4-element ordering to the pair-scoped elim<->+2pp swap (the cap_150k/+1pp legs report-only, anchor_ordering_internally_inconsistent_under_certified_swap). hh_size_share.2 stays gated (candidate 3 clears it 5/5). Zero threshold movement; gates.yaml UNTOUCHED -- the flip is a separate ratified PR. Evidence: W1 forensics 3 (runs/gate_w1_forensics3_v1.json, registration 4959668253, grading 4960691167). Q10 -- the cap_150k entailment: the same above-wage-base compression correction that certifies the elim<->+2pp swap mechanically lifts cap_150k to rank 2 (deployed A/B 1.694 PV), so the 4-element ordering is internally inconsistent as a representative-frame transport target (elim>+2pp <=> A/B>0.1613 vs the Smith-adjacency window A/B in (0.1613, 0.184), overshot ~10x); the above-cap share owns 0.9925 of the gap; contract-permitted lever restoration enumerates EMPTY; the pair-scoped C2 is anchor-supported (Smith 2015 p.3: elim +21yr > +2pp +18yr) and realised 3/3 across c1/c2/c3. Q11 -- the hh_size levers are exhausted: the roster- ceiling probe reaches terminal coresidence only 0.111 (capped by the certified parental-exit hazard, not the seed) and does not close size-1; three candidates deep, all named levers deployed; the residual needs a non-entry-state household model. Both surfaces were pre-named in the ratified amendment-2 series forecast before c3 ran. The whittling question, head-on (section 3): the post-amendment 44-cell surface (43 family-A joints + the pair-scoped C2) would be PASSABLE by the already-committed candidate-3 model -- c3 lands 43/43 of the surviving family-A cells in-band on all five seeds and realised the elim<->+2pp swap -- a first for this gate. The demotion boundary is the permitted-lever line applied prospectively (pre-named before c3 ran), not a post-hoc fit to c3; candidates 1 (PR #162), 2 (PR #167), and 3 (PR #176) all STAND FAIL (no self-rescue; c3's committed FAIL was scored under the current contract). A candidate-4 registration of the unchanged c3 model follows the flip, subject to the standing independent bit-exact reproduction addendum. OC recompute (frozen floor sigmas): family A 47 -> 43 gated, p_seed 0.9344 -> 0.9403, p_gate 0.9623 -> 0.9684; the gate OC moves from 0.9623 x I(C2) to 0.9684 x I(pair-swap) on the residual 44-cell surface (48 -> 44 gated, 84 -> 88 report-only; cube shape [20,47,5] -> [20,43,5]). Machine-checkable ledger + section-7 flip text string-bound to tests/test_gate_w1_amendment3_proposal.py (23 artifact-tier tests binding every load-bearing figure -- forensics-3 shares/mirror/ceiling, per-cell pass records from all three cubes, OC before/after, partition roll-ups; mutation battery >=6 incl. the blanket-impossibility-reintroduction and whittling-section-deletion guards). Tier manifest recount: artifact 949 -> 972. gates.yaml byte-identical to origin/master. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…d C2 Proposal (DRAFT) to demote the four family-A household-size cells hh_size_share.{1,3,4,5plus} to report-only (uniform machine reason no_permitted_entry_state_lever_reaches_cell) and re-scope the family-C gated binary C2 from the full 4-element ordering to the pair-scoped elim<->+2pp swap (the cap_150k/+1pp legs report-only, anchor_ordering_internally_inconsistent_under_certified_swap). hh_size_share.2 stays gated (candidate 3 clears it 5/5). Zero threshold movement; gates.yaml UNTOUCHED -- the flip is a separate ratified PR. Evidence: W1 forensics 3 (runs/gate_w1_forensics3_v1.json, registration 4959668253, grading 4960691167). Q10 -- the cap_150k entailment: the same above-wage-base compression correction that certifies the elim<->+2pp swap mechanically lifts cap_150k to rank 2 (deployed A/B 1.694 PV), so the 4-element ordering is internally inconsistent as a representative-frame transport target (elim>+2pp <=> A/B>0.1613 vs the Smith-adjacency window A/B in (0.1613, 0.184), overshot ~10x); the above-cap share owns 0.9925 of the gap; contract-permitted lever restoration enumerates EMPTY; the pair-scoped C2 is anchor-supported (Smith 2015 p.3: elim +21yr > +2pp +18yr) and realised 3/3 across c1/c2/c3. Q11 -- the hh_size levers are exhausted: the roster- ceiling probe reaches terminal coresidence only 0.111 (capped by the certified parental-exit hazard, not the seed) and does not close size-1; three candidates deep, all named levers deployed; the residual needs a non-entry-state household model. Both surfaces were pre-named in the ratified amendment-2 series forecast before c3 ran. The whittling question, head-on (section 3): the post-amendment 44-cell surface (43 family-A joints + the pair-scoped C2) would be PASSABLE by the already-committed candidate-3 model -- c3 lands 43/43 of the surviving family-A cells in-band on all five seeds and realised the elim<->+2pp swap -- a first for this gate. The demotion boundary is the permitted-lever line applied prospectively (pre-named before c3 ran), not a post-hoc fit to c3; candidates 1 (PR #162), 2 (PR #167), and 3 (PR #176) all STAND FAIL (no self-rescue; c3's committed FAIL was scored under the current contract). A candidate-4 registration of the unchanged c3 model follows the flip, subject to the standing independent bit-exact reproduction addendum. OC recompute (frozen floor sigmas): family A 47 -> 43 gated, p_seed 0.9344 -> 0.9403, p_gate 0.9623 -> 0.9684; the gate OC moves from 0.9623 x I(C2) to 0.9684 x I(pair-swap) on the residual 44-cell surface (48 -> 44 gated, 84 -> 88 report-only; cube shape [20,47,5] -> [20,43,5]). Machine-checkable ledger + section-7 flip text string-bound to tests/test_gate_w1_amendment3_proposal.py (23 artifact-tier tests binding every load-bearing figure -- forensics-3 shares/mirror/ceiling, per-cell pass records from all three cubes, OC before/after, partition roll-ups; mutation battery >=6 incl. the blanket-impossibility-reintroduction and whittling-section-deletion guards). Tier manifest recount: artifact 965 -> 988 (rebased onto the gate_m6 lock, #177). gates.yaml byte-identical to origin/master. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…d after five-round ceremony) (#180) * gate_w1 amendment 3 (ceremony pending): hh_size demotion + pair-scoped C2 Proposal (DRAFT) to demote the four family-A household-size cells hh_size_share.{1,3,4,5plus} to report-only (uniform machine reason no_permitted_entry_state_lever_reaches_cell) and re-scope the family-C gated binary C2 from the full 4-element ordering to the pair-scoped elim<->+2pp swap (the cap_150k/+1pp legs report-only, anchor_ordering_internally_inconsistent_under_certified_swap). hh_size_share.2 stays gated (candidate 3 clears it 5/5). Zero threshold movement; gates.yaml UNTOUCHED -- the flip is a separate ratified PR. Evidence: W1 forensics 3 (runs/gate_w1_forensics3_v1.json, registration 4959668253, grading 4960691167). Q10 -- the cap_150k entailment: the same above-wage-base compression correction that certifies the elim<->+2pp swap mechanically lifts cap_150k to rank 2 (deployed A/B 1.694 PV), so the 4-element ordering is internally inconsistent as a representative-frame transport target (elim>+2pp <=> A/B>0.1613 vs the Smith-adjacency window A/B in (0.1613, 0.184), overshot ~10x); the above-cap share owns 0.9925 of the gap; contract-permitted lever restoration enumerates EMPTY; the pair-scoped C2 is anchor-supported (Smith 2015 p.3: elim +21yr > +2pp +18yr) and realised 3/3 across c1/c2/c3. Q11 -- the hh_size levers are exhausted: the roster- ceiling probe reaches terminal coresidence only 0.111 (capped by the certified parental-exit hazard, not the seed) and does not close size-1; three candidates deep, all named levers deployed; the residual needs a non-entry-state household model. Both surfaces were pre-named in the ratified amendment-2 series forecast before c3 ran. The whittling question, head-on (section 3): the post-amendment 44-cell surface (43 family-A joints + the pair-scoped C2) would be PASSABLE by the already-committed candidate-3 model -- c3 lands 43/43 of the surviving family-A cells in-band on all five seeds and realised the elim<->+2pp swap -- a first for this gate. The demotion boundary is the permitted-lever line applied prospectively (pre-named before c3 ran), not a post-hoc fit to c3; candidates 1 (PR #162), 2 (PR #167), and 3 (PR #176) all STAND FAIL (no self-rescue; c3's committed FAIL was scored under the current contract). A candidate-4 registration of the unchanged c3 model follows the flip, subject to the standing independent bit-exact reproduction addendum. OC recompute (frozen floor sigmas): family A 47 -> 43 gated, p_seed 0.9344 -> 0.9403, p_gate 0.9623 -> 0.9684; the gate OC moves from 0.9623 x I(C2) to 0.9684 x I(pair-swap) on the residual 44-cell surface (48 -> 44 gated, 84 -> 88 report-only; cube shape [20,47,5] -> [20,43,5]). Machine-checkable ledger + section-7 flip text string-bound to tests/test_gate_w1_amendment3_proposal.py (23 artifact-tier tests binding every load-bearing figure -- forensics-3 shares/mirror/ceiling, per-cell pass records from all three cubes, OC before/after, partition roll-ups; mutation battery >=6 incl. the blanket-impossibility-reintroduction and whittling-section-deletion guards). Tier manifest recount: artifact 965 -> 988 (rebased onto the gate_m6 lock, #177). gates.yaml byte-identical to origin/master. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * gate_w1 amendment 3 fixes round 1: scope the C2 impossibility to the pinned frame Addresses the adversarial referee round (PR #180 comment 4962052314, verdict AMEND THE AMENDMENT -- recommendation survives on every recomputed probe P1/P2/P4/P5/P6). Two grounds-precision findings; the demote-quad + pair-scope-C2 recommendation is UNCHANGED. F1 (primary -- the a2 defect class re-instantiated in the C2 argument): sections 2a/1/4b-c-c/8/10 asserted as a UNIVERSAL that "a frame cannot both realise the certified swap and reproduce Smith's cap_150k rank / the 4-element ordering is internally inconsistent as a representative-frame transport target". False as a universal: the order-survival window is f-conditional (0.1613, 0.0806/f), widening to ~(0.1613, 0.37) at the near-band fraction f~=0.22, and forensics-3's own order_at_f_deployed_0p22 = [elimination, payroll_plus_2pp, payroll_plus_1pp, cap_150k] exhibits a representative frame (A/B~=0.22, f~=0.22) that realises the swap AND reproduces the full 4-element Smith order (tau 1.0). Fix: scope the impossibility to the PINNED certified frame throughout, and lead with the two dispositive grounds that carry it -- (1) the pinned/deployed frame measures cap at rank 2 bit-identically (A_band/B 0.373, cap 16.7yr, reproduced c1/c2/c3), so the pinned transport target VIOLATES the full order while realising the swap; (2) no published anchor supports the 4-element order on a representative frame. The f-conditional window arithmetic is retained as analysis with the caveat stated plainly (a swap-consistent full-order frame exists at other f -- forensics-3 order_at_f_deployed_0p22). The machine reason anchor_ordering_internally_inconsistent_under_certified_swap is retained (the referee ruled its under_certified_swap qualifier already scopes correctly) with a new pinned-frame ledger definition (cap_150k_impossibility block, scope pinned_certified_frame). F2 (minor): the fertility lever's "support floor" now reads "support ceiling"/maxed consistently (it is maxed at the 15-entry certified support, not at a floor); the lever-exhaustion claim is unaffected. Bindings: +3 tests (26 total), incl. the pinned-frame-scope + f-conditional refutation binding, a universal-C2-phrasing mutation guard, and a fertility- terminology guard. Full mutation battery re-run: all previously-caught still caught + the two new guards (16 distinct). Tier recount artifact 988 -> 991 (+3). Suite green (unit+artifact 1217 passed / 41 skipped, 0 failed). gates.yaml UNTOUCHED (byte- identical to origin/master). Ceremony record updated (section 8 referee_round + section 11 checklist). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * gate_w1 amendment 3 fixes round 2: scope the §2a restoration clause to the pinned frame + extend the universal-phrasing guard Responds to verification comment 4962578395 (verdict FURTHER FIXES x2 -- both grounds-precision text edits, recommendation untouched). V1 (PR-body squash message; a GitHub edit, not in this commit): the Evidence Q10 bullet carried the round-1 universal ("internally inconsistent as a representative-frame transport target ... overshot ~10x", f-fixed window, no caveat) verbatim; rewritten to the pinned-frame grounds (cap at rank 2, A_band/B 0.373 bit-identical; no published representative-frame anchor) with the f-conditional window (0.1613, 0.0806/f) kept as analysis and the order_at_f_deployed_0p22 counterexample stated -- mirrors the section-8 history entry. V2 (this commit -- section-2a restoration clause): "no unimodal earnings distribution makes it small enough while landing A/B in the window" was an unscoped frame-level universal that the fix's own caveat ten lines above contradicts. Rescoped to the pinned frame's binding constraint: the pinned A/B 1.694 would need f < 0.0476 (= (0.01/r)/AB = 0.01/d_bal_elim, recomputed from the artifact) to land in-window, far below the measured deployed f 0.220; the artifact's "small enough (<0.5)" gloss and the "no contract-permitted lever" conjunct are restored, and the only construction that could move a pinned A/B that far (tail back-solving) is named as the prohibited identity-in-disguise. Bindings/guard: - Ledger field pinned_A_over_B_needs_f_below_round4 0.0476, recomputed from the forensics-3 revenue components in test_q10_impossibility_scoped_to_pinned_frame_not_universal. - BANNED_UNIVERSAL_C2_PHRASINGS extended past the round-1 slots to the earnings-distribution / representative-frame synonym class -- the verifier's M4 ("no earnings distribution of any shape ...") and M4b ("no representative frame whatsoever ...") mutations are now caught, along with the section-2a slot's "no unimodal earnings distribution makes it small enough". - The caveat-existence binding tightened from OR to AND (verifier soft spot 2: M2b-style deletion of the existence sentence is now caught). 26 module bindings (unchanged function count; +1 bound figure, f<0.0476), tier artifact 991 unchanged (no new collected item), always-runnable suite 1217 passed / 41 skipped / 0 failed, black -l 79 + ruff clean. gates.yaml byte-identical to origin/master; only the doc + test changed. Stays DRAFT. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

W1 candidate 1: first full transport deployment (registration 4950931…

8f77ca6

…131) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

MaxGhenis force-pushed the w1-candidate1 branch from 18b9ed7 to 8f77ca6 Compare July 12, 2026 12:38

vercel Bot deployed to Preview July 12, 2026 12:38 View deployment

MaxGhenis merged commit e53b58c into master Jul 12, 2026
7 checks passed

This was referenced Jul 12, 2026

Candidate 2 design: latent-permanent conditioned chained QRF (research memo) #42

Open

gate_w1 amendment 1 (draft — ceremony pending): family-B DI bands #164

Merged

MaxGhenis mentioned this pull request Jul 12, 2026

Paper: the transport ladder, the disability gate, and a gate amended by its own evidence #166

Merged

MaxGhenis mentioned this pull request Jul 13, 2026

gate_w1 amendment 2 (draft — ceremony pending): family-A concept cells #169

Merged

8 tasks

MaxGhenis mentioned this pull request Jul 13, 2026

gate_w1 amendment 3: demote the hh_size quad + pair-scope C2 (ratified after five-round ceremony) #180

Merged

Navigation Menu

Uh oh!

W1 candidate 1: first full transport deployment (registration 4950931131) - #162

W1 candidate 1: first full transport deployment (registration 4950931131)#162
MaxGhenis merged 1 commit into
masterfrom
w1-candidate1

MaxGhenis commented Jul 12, 2026

Uh oh!

vercel Bot commented Jul 12, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

MaxGhenis commented Jul 12, 2026

Verdict: FAIL (0 / 5 seeds)

Per-seed verdict table (family A, 53 gated cells)

The fingerprint reversal results (family C — the highest-value read)

Per-family decomposition

Spec resolutions (documented; the locked text does not pin them)

Reproduction

Tests

Uh oh!

vercel Bot commented Jul 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

vercel Bot commented Jul 12, 2026 •

edited

Loading