gate_w1 candidate 3: one-shot scored run (registered 4959017270) by MaxGhenis · Pull Request #176 · PolicyEngine/populace-dynamics

MaxGhenis · 2026-07-13T15:23:47Z

Verdict: gate FAIL

One-shot scored run of W1 candidate 3 against the live amended gate_w1 surface (48 gated = 47 family-A + C2), partition bound from gates.yaml at run time. Published regardless of verdict (no_self_rescue; write_new one-shot, refuses overwrite). Candidate had never touched the pinned frame before this run.

Registration: issue #42, comment 4959017270 (posted after the amendment-2 flip d250d4b; scored surface fixed before the run).

Artifact: runs/gate_w1_candidate3_v1.json (git blob 3265098e94a979a497a59a8cfc85615435601bf3) + env sidecar runs/gate_w1_candidate3_v1.json.env.json.

Gate outcome

Component	Rule	Result
Family A	≥ 4 / 5 seeds fully in-band (47/47)	FAIL — 0 / 5 seeds pass
Family C (gated: C2)	every gated fingerprint reverses to anchor	FAIL — C2 not reversed
Family B	report-only, contributes nothing	0 / 10 within tol (does not gate)
Gate	Family A AND Family C	FAIL

Family A — per seed (of 47 gated cells)

Seed	In-band	Seed pass	Failing cells
0	43/47	✗	`hh_size_share.{1, 3, 4, 5plus}`
1	44/47	✗	`hh_size_share.{1, 3, 5plus}`
2	43/47	✗	`hh_size_share.{1, 3, 4, 5plus}`
3	43/47	✗	`hh_size_share.{1, 3, 4, 5plus}`
4	43/47	✗	`hh_size_share.{1, 3, 4, 5plus}`

Fail frequency across seeds: hh_size_share.1, hh_size_share.3, hh_size_share.5plus fail 5/5; hh_size_share.4 fails 4/5 (clears on seed 1). hh_size_share.2 (size-2) clears on every seed. No cell outside the four hh_size cells ever fails — the other 43 gated cells are in-band on all 5 seeds (including every 25-34 marital_share cell).

Per-draw dispersion (conformance): max_across_draw_sd = 0.351, max_per_draw_abs_ln = 0.938, regenerated_surface = true, identity_candidate = false — non-zero across-draw dispersion confirms the candidate regenerates every scored column (the prohibited identity map is excluded).

Family C — compression fingerprints

Fingerprint	Gated	Reversed to anchor	τ vs required	Required swap realised
C2 `elimination_plus2pp`	yes	False	0.333	True
C1 `ppi_nra`	no (report-only)	False	0.667	False

C2 (tau + adjacency): the required adjacent swap pair [payroll_plus_2pp ↔ elimination] is realised (elimination moves to the top), but the full order does not reverse to the representative anchor — cap_150k lands in position 2:

deployed: [elimination, cap_150k, payroll_plus_2pp, payroll_plus_1pp]
required: [elimination, payroll_plus_2pp, payroll_plus_1pp, cap_150k]

so kendall_tau_vs_required = 0.333 and reversed_to_anchor = False. (kendall_tau_vs_psid = 0.0.)

Report-only families (published, do not gate)

C1 (ppi_nra): deployed order equals the PSID frame order (no swap; kendall_tau_vs_psid = 1.0), reversed_to_anchor = False, τ vs required 0.667.
Family B (10 M4/DI margins — di_prevalence.{under30,30-34,35-39,40-44,45-49,50-54,55-59,60-fra} + claim_age.disability_conversion|{female,male}): 0/10 within tolerance.
18-24 pair / 65+ quad (demoted from gating by amendment 2 — earnings_participation.18-24|{f,m} and marital_share.married.65+|{f,m} + coresident_spouse.65+|{f,m}): the run binds the scored partition to the 47 gated cells only, so these demoted cells are not evaluated in this artifact.

Registered disclosures (restated)

Lever-1 partial self-reference. The entry anchor reads the same 25-34 band whose marital-share cells remain gated. It is the Q6-permitted entry channel, but for those cells the run is partially anchored to the scored quantity's own band average: the anchor pins an age-25 boundary condition, while the gated cells score the full 25-34 band after up-to-nine-years evolution under the locked hazards, plus every other band and view. Readers should weight 25-34 marital passes accordingly. (Outcome: all 25-34 marital_share cells were in-band on all 5 seeds; none are among the failures.)
Effective search size. Per the standing addendum (comment 4948637741), this is the third scored candidate in the W1 family, following two forensics rounds and two ratified amendments. The family-level search is the honest unit; per-candidate framing would underprice it. Every prior run in the family is registered and publicly graded (registry 51/51 at registration).
The surface is not known passable (amendment-2 candor, ratified c50f8bb). The four hh_size cells (sizes 1/3/4/5+) are pre-named amendment-3 risks — Q7 was graded REFUTED for closability by the two entry-state levers alone ("not closable by these two entry-state levers alone"); lever 3's co-design targets exactly that coresidence-composition residual and is the least-proven of the three levers.

Forecast vs outcome

Registered forecast	Outcome
p(gate PASS) ≈ 0.15	FAIL (in the ~85% modal mass)
Modal fail = the four pre-named hh_size cells (sizes 1/3/4/5+)	EXACT MATCH — the failure surface is precisely `hh_size_share.{1, 3, 4, 5plus}`; nothing beyond it; size-2 clears exactly as Q7's arithmetic predicted
C2 reverses, p ≈ 0.9	Did NOT reverse — the ~10% tail realised (the required swap pair fired, but `cap_150k` mis-ordered kept τ at 0.333); this is the forecast's Modal-residual-#2 direction

Amendment-3 condition. The registration named the amendment-3 evidence as "the failure surface is EXACTLY those four cells (with the rest of family A in-band at ≥4 seeds and C2 reversed)." Two of three clauses hold — the family-A surface is exactly the four cells and the other 43 cells are in-band at 5/5 seeds — but C2 did not reverse, so the full trigger is not met. The family-A surface is not "materially beyond" the four cells (which would force new forensics on family A); C2's non-reversal is the low-probability branch the forecast itself named and is the item requiring adjudication.

Env replicates candidate 2 exactly: frame exported in the policyengine.py .venv (certified frame sha256 c2065b64… verified, 166302 persons / 57240 households); scored phases in .venv-gate (py 3.14.4 / numpy 2.5.1 / pandas 3.0.3 / sklearn 1.8.0 / scipy 1.18.0 — matches the c2 sidecar byte-for-byte), populace_dynamics from this worktree.

Draft — grading and the registry update (→ 52/52) are the orchestrator's act; not marking ready.

🤖 Generated with Claude Code

…te FAIL Scored the registered candidate-3 one-shot against the live amended gate_w1 surface (48 gated = 47 family-A + C2), bound from gates.yaml at run time. Published regardless of verdict (no_self_rescue; write_new). Verdict: gate FAIL. - Family A: 0/5 seeds pass (in-band 43/44/43/43/43 of 47). The ONLY failing cells are the four pre-named hh_size_share cells (sizes 1/3/4/5+): sizes 1/3/5plus fail 5/5 seeds, size 4 fails 4/5 (clears on seed 1). Size-2 clears. Failure surface is EXACTLY the four amendment-3 cells — nothing beyond them. - Family C: gated C2 (elimination_plus2pp) does NOT reverse to anchor (kendall_tau_vs_required 0.333; the required elimination<->+2pp swap pair IS realised, but cap_150k lands out of order). C1 report-only, not reversed (tau 0.667). - Family B: report-only, 0/10 within tolerance (gates nothing). Env replicates candidate 2 exactly: frame export in policyengine.py .venv (certified frame sha256 c2065b64 verified, 166302 persons); scored phases in .venv-gate (py 3.14.4 / numpy 2.5.1 / pandas 3.0.3 / sklearn 1.8.0 / scipy 1.18.0), populace_dynamics from this worktree. Artifact + env sidecar written via write_new (one-shot, refuses overwrite). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

vercel · 2026-07-13T15:23:54Z

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 3:23pm

…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>

MaxGhenis marked this pull request as ready for review July 13, 2026 15:29

MaxGhenis merged commit 64ebf97 into master Jul 13, 2026
7 checks passed

This was referenced Jul 13, 2026

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

Open

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

Merged

MaxGhenis mentioned this pull request Jul 14, 2026

gate_w1 candidate 4: one-shot on the amendment-3 surface (registered 4964126356) #184

Draft

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

gate_w1 candidate 3: one-shot scored run (registered 4959017270)#176

gate_w1 candidate 3: one-shot scored run (registered 4959017270)#176
MaxGhenis merged 1 commit into
masterfrom
w1-candidate3-run

MaxGhenis commented Jul 13, 2026

Uh oh!

vercel Bot commented Jul 13, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

MaxGhenis commented Jul 13, 2026

Verdict: gate FAIL

Gate outcome

Family A — per seed (of 47 gated cells)

Family C — compression fingerprints

Report-only families (published, do not gate)

Registered disclosures (restated)

Forecast vs outcome

Uh oh!

vercel Bot commented Jul 13, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant