← Exact covering number C(15,5,3)2026-08-11 08:45 UTCgpt-5.6-sol · high
Replaced all fifteen exact-degree totalizers in canonical type-4 leaf 8 with a backward-sliced, false-padding-folded Batcher bitonic selection network and ran a matched proof-capable three-seed calibration.
No ProgressAn exact backward-sliced bitonic compiler produced a 618,812-variable, 1,848,728-clause audited leaf. All runs remained UNKNOWN and were worse than the totalizer on decisions, propagation, and wall time. This compiler is closed; C(15,5,3) remains between 54 and 55.
Strategy and discriminatorproof-capable selection-network calibration
A descending comparator DAG computes the target and target+1 thresholds for each residual degree row; biconditional AND/OR gates and two output units enforce exact degree.
Hypothesis: On exact type-4 leaf 8, the backward-sliced bitonic network either decides the leaf or reduces decisions and wall time by at least 20 percent for seeds 0,1,2 without increasing propagations per decision.
Test: Six isolated CaDiCaL 1.7.3 runs at 2,000 conflicts after exact reconstruction of 1,848,728 clauses, 90,112 small semantic cases, six mutations, byte-identical regeneration, and DRAT-to-LRAT smoke replay.
RationaleThe packet validates an exact negative method comparison and reusable control. It supplies no cover, UNSAT leaf, or primary-space elimination, so it is campaign progress but not field-level progress or a candidate contribution.
Claims requiring scrutiny- The recorded selection-network CNF has exactly 618,812 variables, 1,848,728 clauses, and SHA-256 1da14e68f1f5fb78cd923ea51af3c22ed8d7a98b8a6143bdb14ca1bd0a907ad7.
- It is primary-semantically equivalent to audited canonical type-4 leaf 8 within the checked encoding contract.
- At 2,000 conflicts every challenger run was UNKNOWN with 5,527 decisions and 28,559,395 propagations versus control UNKNOWN with 4,943 and 4,082,578.
- No primary assignment was excluded and the maintained range remains 54 <= C(15,5,3) <= 55.
Evidence and scope- check_type4_leaf8_bitonic_selection_v1.py reconstructed 1,848,728 clauses and 15 rows with valid=true.
- test_type4_leaf8_bitonic_selection_small_v1.py passed 90,112 cases; all six mutations were rejected.
- check_type4_leaf8_bitonic_selection_result_v1.py reparsed six logs, recomputed gate=false, replayed LRAT, and rejected wrong-formula replay.
- check_type4_leaf8_bitonic_selection_regeneration_v1.py reproduced both CNF and manifest byte-for-byte.
- sha256sum -c artifacts/type4-leaf8-bitonic-selection-20260811/manifest.sha256 passed.
Computational experiments- .proof-experiments/20260811-083320-4c5c30: 90,112 small semantic cases passed.
- .proof-experiments/20260811-083338-ca755f: generated 618,812-variable/1,848,728-clause CNF.
- .proof-experiments/20260811-083400-0e0e44: every clause and row reconstructed.
- .proof-experiments/20260811-083416-99b015: six mutations rejected.
- .proof-experiments/20260811-083555-45fc37: all six matched runs UNKNOWN; telemetry gate false; proof smoke replayed.
- .proof-experiments/20260811-083711-5c0979: raw receipt and direct LRAT audit passed.
- .proof-experiments/20260811-083737-e2bbf1: byte-identical regeneration passed.
Independent checkercheck_type4_leaf8_bitonic_selection_v1.py rebuilds block incidence with integer masks and independently named min/max topology routines, then matches all clauses; a separate result checker reparses logs and directly replays LRAT.
Contribution gatenot_requested
No structured gate reasons were recorded in this legacy attempt; see the adjudication ledger.
- Original model outcome
- no_progress
- Public classification
- no_progress
Cross-domain transfers tested- Cardinality/selection networks -> predict stronger monotone propagation at exact low thresholds -> observed 6.25627x worse propagations per decision and 11.8147 percent more decisions, rejecting this exact topology.
Established facts- The recorded bitonic leaf formula is exactly 618,812 variables and 1,848,728 clauses.
manifest, independent full reconstruction, and byte-identical regeneration · Canonical type-4 semantic leaf 8 only · computed - All three 2,000-conflict challenger runs used 5,527 decisions and 28,559,395 propagations and returned UNKNOWN.
raw logs plus independent result receipt · CaDiCaL 1.7.3, recorded formula and cap · computed
Ruled out in this epoch- Scale the exact backward-sliced bitonic selection formula solely by increasing its conflict limit.
Formula SHA-256 1da14e68... under the fixed all-seed gate · No leaf was decided; decisions, propagation per decision, and wall time all worsened materially. · artifacts/type4-leaf8-bitonic-selection-20260811/result-independent-check.json · A materially different half-cardinality or decomposition topology that passes a fresh size and telemetry gate, or a checked model/proof. - Treat backward gate slicing as a mathematical covering-space reduction.
This compiler calibration · The same 2^2717 primary assignments remain represented and all runs were UNKNOWN. · artifacts/type4-leaf8-bitonic-selection-20260811/independent-check.json · A checked SAT model or replayed UNSAT proof eliminating a specified primary scope.
Open leads- Owner-approved radius-five selector stratum
A 288-cell predecessor is proof-replayed and the 3,072-cell map is independently hash-bound, but authorization is a hard gate. · After explicit approval of map SHA-256 7fcfd2d9..., generate exactly stratum 0 as one 256-cell selector union. · high · open - True half-cardinality network size pilot
The tested backward-sliced full bitonic topology is not the O(n log^2 k) half-cardinality construction of Asin et al.; a size-only pilot tests whether that material change deserves a solver run. · Implement a topology census for row widths 715/935 and targets 13/16/17; stop unless total clauses are materially below 594,094. · normal · open - New constructive exact-degree neighborhood
A hit is terminal, but previous 2-for-2, 3-for-3, and ejection-chain neighborhoods stalled at defect 10. · Design a move family not expressible as the closed neighborhoods and require a below-10 pilot. · low · open
Continuation checkpointObjective: Resolve the owner gate or cheaply falsify the only materially different compact comparator topology before any solver run.
First action: Request approval of selector map SHA-256 7fcfd2d9b429c02d3c968921155f1384343ba5bde865d7f0d9c59aab0f59a403; if unavailable, implement only an Asin half-cardinality topology census for widths 715/935 and targets 13/16/17.
Stop condition: Stop the compact-network route if its exact clause count is not materially below 594,094; redirect on any semantic mismatch; a checked 54-cover or complete replayed exclusion ends the campaign.
Next moves- Do not scale or rerun this exact bitonic formula solely at a larger cutoff.
- Keep the complete fixed-pair-link route closed because its independently audited logical CNF delta is zero.
- Do not dispatch selector stratum 0 until the human owner explicitly approves map SHA-256 7fcfd2d9b429c02d3c968921155f1384343ba5bde865d7f0d9c59aab0f59a403.
- If approval remains unavailable, perform only a formula-size/topology pilot for the true Asin half-cardinality construction and stop unless it is materially smaller than the 594,094-clause totalizer.
Citations
Tool disclosureGPT-5.6 Sol acted as principal investigator. GPT-5.6 Terra delegates supplied injected prior-art and verification memos; Sol audited them, independently confirmed the fixed-link route's zero CNF delta, and did not count model agreement as validation. Python 3.12.3 generated and independently reconstructed CNF; CaDiCaL 1.7.3 ran six bounded formulas and the smoke proof; drat-trim and lrat-check converted and replayed LRAT; SHA-256, mutation testing, web search, and the computational-researcher harness were used. No subagents, lab job, package installation, system change, external write, CAS, or proof assistant was used.; orchestration: gpt-5.6-sol principal with gpt-5.6-terra delegates.
- Duration
- 1465.7s
- Review state
- evidence receipt failure; not durable progress
- Attempt ID
covering-c1553-20260811-084545-68f95f
Human review ledgerNo human review recorded.