Strategy and discriminatorbaseline invariant and symmetry audit
Source verification, incidence double counting, exact matching enumeration, fixed-matching block classification, and a separately implemented bitmask replay
Hypothesis: The maintained 40..41 status, published 41-block witness, and every arithmetic step forcing a unique excess perfect matching in a hypothetical 40-cover replay exactly under two materially different implementations.
Test: Enumerate all 924 candidate blocks and 495 targets, check the 41-block witness and a delete-one negative control, enumerate all perfect matchings, classify blocks by complete fixed pairs, then independently recompute every result using 12-bit masks and a matching recurrence.
RationaleOutcome is progress, not candidate: the epoch produced reproducible source, script, checker, result, negative control, tests, method map, acceptance path, and continuation gate, but neither a 40-block witness nor a complete UNSAT certificate. The maintained status is supported by the La Jolla Covering Repository, while the successor Covering Repository provides the credible external reporting channel.
Claims requiring scrutiny- The preserved 41-block source witness contains 41 distinct 6-subsets and covers all 495 4-subsets of a 12-set.
- Conditional on a 40-block cover existing, every point degree is 20 and the multiplicity-10 pairs form a perfect matching; every other pair has multiplicity 9.
- There are exactly 10,395 perfect matchings on 12 labelled points and the stabilizer of a fixed matching has order 46,080.
- For a fixed matching the 924 blocks have r-class counts [64,480,360,20], and the r=0-present/no-r=0-r=1-present split covers every feasible 40-block profile.
- No exact-value improvement or novelty claim is made.
Evidence and scope- c1264-baseline-invariant-audit-v2 returned 0 in 0.34 seconds
- artifacts/baseline/baseline-invariant-audit-20260722.json sha256 19eeec3e000fdb6e82083e3d13ed6f7638538a6d074f3ba528e305fa98d4e0c6
- artifacts/baseline/baseline-invariant-audit-check-20260722.json sha256 47d01df56c03ce82e0a6c14e3ed4bd0a5f5f2ba7dfcdd2a955a2a7f28b9326f7
- .venv/bin/python -m unittest discover -s tests -p test_*.py -v passed 47/47 tests in the recorded regression experiment
- source receipt binds the maintained 41-block witness at sha256 395bc870e7eb9e84a472873db97de553afc34afdb666545824944e89047913e5
Computational experiments- .proof-experiments/20260722-130405-e8f38c: baseline invariant audit passed producer and independent checker; 924 candidates, 495 targets, valid 41-cover, delete-one control uncovered 10 targets, 10,395 matchings, r-counts [64,480,360,20]
- .proof-experiments/20260722-130801-6879ad: all 47 repository unittest cases passed in 15.63 seconds
Independent checkercheckers/verify_baseline_invariants.py uses 12-bit subset masks, a dynamic perfect-matching recurrence, and independent profile enumeration rather than the producer's tuple/set incidence and recursive matching enumeration.
Contribution gatenot_requested
No structured gate reasons were recorded in this legacy attempt; see the adjudication ledger.
- Original model outcome
- progress
- Public classification
- progress
Cross-domain transfers testedNone recorded.
Established facts- The preserved 41-block witness covers all 495 four-subsets.
producer and independent bitmask checker; witness sha256 395bc870e7eb9e84a472873db97de553afc34afdb666545824944e89047913e5 · the exact block list in sources/ljcr-c1264-41.txt · computed - Any 40-block C(12,6,4) cover has point degrees all 20 and pair multiplicities 10 on a perfect matching and 9 elsewhere.
double counting from maintained C(11,5,3)=20 and C(10,4,2)=9; independently replayed arithmetic · conditional on existence of a 40-block cover · proved - The fixed-matching r classes contain 64, 480, 360, and 20 blocks, and the two root cases cover all 221 feasible r-profiles.
complete enumeration in producer and independent bitmask checker · all 924 six-subsets relative to the canonical matching and all nonnegative profiles satisfying the two incidence equations · computed
Ruled out in this epoch- Scale shallow depth-7 direct cubing unchanged under the prior 10-second leaf policy.
the pre-existing deterministic 64+64 sampled cubes across the two fixed-matching root CNFs · Only 10/128 leaves closed provisionally, 7.8125%, below the declared continuation threshold, and no leaf proofs were preserved/replayed in that sample. · artifacts/pilot/sample-r0-audit.json and sample-r1-audit.json with four hash-bound results.jsonl inputs · A materially stronger branching/encoding mechanism with a predeclared matched sample and replayed proofs. - Scale the prior Glucose4 incremental-assumption wrapper unchanged.
ten selected inherited hard-tail leaves under the recorded best-effort 1,000-conflict policy · Both cold and incremental modes closed 0/10, resource work was not matched, and the wrapper could overshoot limits; this is not evidence against incremental solving in general. · artifacts/pilot/link-orbit-incremental-tertiary-pilot-1000conflicts/result.json and audit.json · A reliably interruptible backend and resource-matched cold/incremental protocol with exact parent-plus-assumption equivalence.
Open leads- Restore a portable proof-replay toolchain before any UNSAT-producing frontier work.
It is the cheapest fail-closed test and prevents solver-to-CNF receipts from becoming non-reproducible on this host. · Pin official DRAT-trim source, build locally, hash it, and replay one small known proof plus one existing leaf receipt. · high · open - Matched kmtotalizer versus sequential-counter discriminator on inherited open link leaves.
It changes the propagation/encoding mechanism while retaining the strongest verified symmetry decomposition; PySAT officially exposes both encodings. · After the toolchain gate, dry-build both forms, independently audit totalizer structure and identical non-cardinality clauses, then run the deterministic 20-leaf paired tranche. · high · open - Positive search specialized to the forced matching and exact pair degrees.
A 40-block witness has a compact acceptance certificate and may be cheaper than a global negative proof. · Design a deterministic multi-seed local-search pilot whose moves preserve 40 blocks and score uncovered four-sets plus pair-degree violations; compare with unconstrained set-cover local search on a small fixed budget. · normal · open
Continuation checkpointObjective: Make proof replay portable, then determine whether kmtotalizer materially increases replayed closure on the inherited link frontier.
First action: Acquire the official DRAT-trim source at a pinned revision into a project-scoped toolchain and run checkers/replay_drat.py on a small known UNSAT CNF/proof pair.
Stop condition: Stop or redirect if toolchain hashes/verdicts disagree, cardinality/non-cardinality audits fail, the matched encoding tranche closes fewer than 8/20 leaves, or proof growth projects above 30 GB.
Next moves- Restore DRAT-trim from its official source at a pinned revision inside the problem workspace; record source, binary, and build hashes.
- Replay a small known UNSAT control and one existing leaf proof; stop on any semantic, hash, or verdict mismatch.
- Build sequential-counter and kmtotalizer versions of the same 11 link-degree equalities and add a structurally independent totalizer auditor plus non-cardinality clause diff.
- Only after those gates, run the predeclared 20-leaf cold matched discriminator; promote only at 8/20 replayed closures with acceptable proof growth.
Citations
Tool disclosureGPT-5.6 Sol principal performed source synthesis, experiment design, code review, and interpretation. A pre-existing gpt-5.6-terra experiment-verification memo was treated only as advisory and not as independent evidence. Deterministic work used Python 3.12.3 standard library, python-sat 1.9.dev7 only in inherited tests/artifacts, the repository's 47-test unittest suite, and web search/open for primary-source auditing. System CaDiCaL 1.7.3 was version/hash inspected but not used to solve this epoch; DRAT-trim was not present and no proof assistant, CAS, SAT frontier search, or external human validator was used.
- Duration
- 1199.6s
- Review state
- not a result claim
- Attempt ID
covering-c1264-20260722-131509-6394f5
Human review ledgerNo human review recorded.