Audit #2751 independently and reconcile it with #2720's T74 centre floor (no κ; no defect found)

research/t74-2751-joint-twelve-audit/TASK.json

@@ -0,0 +1,73 @@
+{
+  "schema_version": 1,
+  "task_id": "independent-moment",
+  "priority_revision": 5,
+  "base_commit": "4cc9d09a124e06dd358fdc1c3d4b2af51dd44649",
+  "original_card": {
+    "id": "independent-moment",
+    "question": "Does one changed supplier or assembly claim lack a source-bound independent check of a specific new implication?",
+    "starting_points": [
+      "certificates/selected-result.json"
+    ],
+    "methods": [
+      "Reconstruct the rank ledger from original operations rather than trusting its aggregate.",
+      "Implement certified moment inequalities through a different exact enclosure method."
+    ],
+    "acceptance_gate": "Demonstrate independence, source binding, positive cases and semantic failure controls. Name the existing checks and the missing coverage. More digits or another equivalent evaluation of an unchanged root is not by itself new verification coverage.",
+    "first_step": {
+      "command": [
+        "python3",
+        "-B",
+        "scripts/research_baseline.py"
+      ],
+      "expected": "PASS_RESEARCH_BASELINE_BINDINGS_ONLY",
+      "inspect": "Record the selected certificate and complex_profile rather than taking a PR headline as input.",
+      "scope": "Positive control / source-bound orientation only; this is not a new construction or full cold replay."
+    },
+    "first_experiment": "List the existing independent checks for the exact selected or candidate sources. Pick a new changed input, missing binding or distinct failure mode; stop with a handoff if the proposed arithmetic check is already covered.",
+    "bottleneck_link": "Independent pricing prevents unsupported frontier claims and can reveal genuine accounting slack.",
+    "allowed_variations": [
+      "Reconstruct either a source ledger or a numeric enclosure with a different method."
+    ],
+    "scope_change_triggers": [
+      "Importing the primary function and calling its output independent verification."
+    ],
+    "overlap_terms": [
+      "independent moment",
+      "rank histogram",
+      "exact enclosure"
+    ]
+  },
+  "scope_status": "refined",
+  "actual_scope": "Independent source-bound verification of open PR #2751 (tekuuu, head 4cfbb0cc; T74 joint twelve-lender recovery, conditional kappa ~0.0017171416), and a computed reconciliation with open PR #2720 (petalas, head 813fae62; the fixed-contract inherited-correction floor). With code independent of the audited PRs, the package checks the following. Both changed words are exact in the free-operator model through cleanup and the target endpoint. All 12 lenders and all 4 pair recoveries are checked. The exact change set and per-word delta 4(H8-H7-H1) are checked against PR2733@2761b52b. PR2720's matrix, residual, continuation and floor are reproduced on its own input words. On PR2751's words, the same projection and PR2720's hypotheses are evaluated as computed booleans from verbatim quotes. The complete profile is rebuilt from 276,390 roles in two implementations, along with stock, the endpoint enclosure, the comparison with PR2733's current head, and the #2766/#2767 profile identity. The structure of the downstream words (#2802, #2818, #2824, #2827, #2833) is checked. It corroborates and extends #2766's registered independent_check.py.",
+  "scope_change_reason": "The loop drew seed 63 (adversarial-audit track). The card is overridden to independent-moment, the closest card for an independent check of a changed supplier claim. PR2751 changes the words, fill and endpoint, and every pending T74 claim read above PR2733 rests on its rank-8 centres. Its only existing independent check is #2766's registered independent_check.py, which covers the word frames and replay. The card's generic question is narrowed to PR2751's finitely checkable implications and to a computed reconciliation with PR2720's floor. PR2751 states that relation itself; this package confirms it independently.",
+  "bottleneck_link": "Every pending T74 claim read at submission above PR2733 rests on the eight centre children being rank eight: #2751, #2766, #2767, #2802, #2818, #2824, #2827 and #2833. Newer claims #2843, #2848 and #2853 build on #2833's construction and were not checked. PR2720 states a floor that applies only while the partner aliases are inherited. An independent check that the gain is real, a computed statement of which PR2720 hypotheses change, and the downstream structure decide what those claims can cite. The selected frontier is unchanged.",
+  "budget_minutes": 120,
+  "stop_rule": "Stop after independently checking every implication of PR2751 listed in the README (words, change set, fill and complete profile, stock, endpoint, kappa) and the PR2720 reconciliation, with semantic negative controls that each name the rule that must reject them. Report the unexecuted native, supplier and outer stages as gaps. No construction, search or optimisation.",
+  "outcome": "completed",
+  "evidence": "verify.py, standard library only, PR code never run (about 3 min, about 1.0 GB here). Both PR2751 words are exact on 639 dirty inputs, forward and dual, through cleanup (4,540 items, equal to the own inverse-reverse) and target endpoint, with the 36 initial corrections emulated exactly. All 12 lenders hold X_r at both handoffs. The 4 pair recoveries restore both sources from ((X_x-X_y)/2, X_x+X_y), with 0 paid climbs. Profile {1:2366,2:646,3:297,4:120,5:26,8:8}; centres 8x[8]; delta {1:-4,7:-4,8:+4} vs PR2733@2761b52b. PR2720 Q det -1/16 is reproduced. PR2751's projection is zero on its rows and has det 1/2 on 21,37,47,110 (never cut-written). PR2720 items changed: h1 (h1a,h1b,h1c), h2a, h3, line30. Kept: h2b, h2c, cut incidences, zero readers, rank-8 producers. The complete profile from 276,390 roles in two implementations equals the claim, with delta 2664(H8-H7-H1); stock 66801; C accepted and C+1e-24 rejected; kappa = C-1e-46-1e-70. Against PR2733@a63c55ff the delta is 2664(H8-H7-H1)+666(2H1-H2), with a C gap of 8.86e-7. #2766 and #2767 profiles are identical. Downstream structure is present in #2802/#2818/#2824/#2827/#2833. 356/356 negative controls are rejected by their named rules (218 distinct); 9 positive controls; 103 checker mutations plus 7 new and a 10-mutation re-run sample killed (local-only). No defect found.",
+  "author": "Sagar Thakrar",
+  "ai_assistance": "Prepared with substantial AI assistance: Claude Code (Anthropic, Claude Opus 5.5) subagents wrote the independent checks and the consolidated package; AI reviewers audited it. No independent human review is claimed.",
+  "open_questions": [
+    {
+      "question": "Are the new K47-K50 charts of PR2751 (changed G and E, same op ids) the same physical operators in the A prelude and at the B handoff, including Clifford phases and monomial lifts?",
+      "why_it_matters": "The free-operator replay treats each warp letter as abstract, so it proves the algebra for any invertible K_op; the native identity is author-checked only. #2766 and #2802 reuse these charts.",
+      "first_step": "Re-implement the 512-column monomial check for ops 47-50 from PR2751's pinned native receipts.",
+      "blocked_by": "Nothing; it needs an independent reading of the native operator encoding."
+    },
+    {
+      "question": "Does PR2720's hypothesis analysis classify #2767's route to 8H8, which moves the partner correction past the cut into B?",
+      "why_it_matters": "#2767 appears to change PR2720's hypothesis 2 (compensation at the start of A) rather than hypothesis 1. Computing the same booleans would show whether both escape routes are understood and whether they can be combined.",
+      "first_step": "Extend audit_reconcile to locate the P corrections in B (via audit_others.parse_inline) and evaluate the PR2720 booleans on #2767's words; PR2720 also excludes 'moving compensation past its boundary'.",
+      "blocked_by": "Nothing."
+    },
+    {
+      "question": "Do PR2751's strict supplier admission and the 63 modern outer inequalities hold when re-derived independently at its 24-digit setting?",
+      "why_it_matters": "kappa = C - 1e-46 - 1e-70 is checked here as arithmetic on the claimed endpoint; the outer stage and native admission are author-checked only.",
+      "first_step": "Restate the inherited ledger formulas and recompute the strict rows from PR2751's SPEC.",
+      "blocked_by": "The inherited outer-ledger formulas must be restated from the PR2462 lineage."
+    }
+  ],
+  "declared_milestone": "audit",
+  "family_expansion_decision": "None proposed. This audit recommends no further word or fill optimisation; its open questions only close named obligations of PR2751's claim and of the PR2720 reconciliation."
+}

research/t74-2751-joint-twelve-audit/VERIFY.json

@@ -0,0 +1,20 @@
+{
+  "schema_version": 1,
+  "status": "ready",
+  "python": "3.12",
+  "requirements": [],
+  "inputs": [],
+  "commands": [
+    {
+      "argv": [
+        "{python}",
+        "-I",
+        "-B",
+        "{package}/verify.py",
+        "--fetch"
+      ],
+      "timeout_seconds": 600,
+      "checks": "Independent verification of open PR #2751 head 4cfbb0cc and reconciliation with open PR #2720 head 813fae62, in standard-library code that reads PR files only as pinned git data (git blob id and sha256) and never imports or runs PR code; GIT_CONFIG* variables are removed before the first git call; --fetch (git >= 2.29, network) writes only the fetched objects and refs/audit/prN. It replays both changed words exactly in the free-operator model (forward and dual, 639 dirty inputs, 36 initial corrections emulated, through cleanup and target endpoint), checks the warp algebra, the frame recount, the cut census, the 12 lender rows and the 4 pair recoveries (each lender and pair controlled), the exact change set and per-word delta 4(H8-H7-H1) against PR2733@2761b52b, PR2720's matrix (det -1/16), omitted-correction error E_P, continuation and cut floor on its own inputs, the same projection on PR2751 (zero on 23,34,50,108; det 1/2 on 21,37,47,110), the echo identities, and one computed boolean per PR2720 hypothesis clause quoted verbatim from its README (PR2751 changes h1, h1a-c, h2a, h3 and the line-30 cut frames; keeps h2b, h2c, the cut incidences, zero readers and rank-8 producers; two premises recorded as not evaluated), the 276390-role fill in two implementations (comparison recorded for every priced fill), the complete profile and delta 2664(H8-H7-H1), the physical stock 66801, the endpoint enclosure and kappa = C-1e-46-1e-70, the comparison with PR2733's head refit a63c55ff (2664(H8-H7-H1)+666(2H1-H2), head profile rejected at PR2751's C), the #2766/#2767 complete-profile identity and #2767's two PROFILE.json files, and, as data only, that the words of #2802, #2818, #2824, #2827 and #2833 keep PR2751's four pair recoveries, ops 47-50 charts and lifted targets; it requires every negative control to be rejected by its named rule (attempted and rejected counted separately), 9 positive controls to pass, and the result to equal RESULT.json. See README.md."
+    }
+  ]
+}