The full span-four theorem remains conditional on finite certificate inputs. This work reduces the remaining weight calculations and supplies checked transport theorems. It does not establish the full weight table.
Candidate data
The original width-four input has 9,088 rows. Its canonical SHA256 is 5594610818abd7f41c76121924259bd1a9a1f86417da91dce2c0f8171bcceb6a. The original input and literal pilot are unchanged.
The orbit producer handles each row at the actual periods seven, eight and nine. It compares the local generator family and its distinguished core, after reducing coordinates modulo the actual period. It uses bijections of their finite supports. It does not collapse boundary points.
| Quantity | Count |
|---|---|
| Original row and period pairs | 27,264 |
| Representative classes | 9,448 |
| Classes with an existing pilot representative | 42 |
| Representatives still needing arithmetic proofs | 9,406 |
| Target weight vectors changed by transport | 6,618 |
Weights are transported from each representative. The target theorem needs a successful weight witness; it does not require the finder's original vector. The candidate artifact records every replacement explicitly. Transport does not prove that each original vector succeeds.
The producer, compressed candidate data, manifest, and independent replay script are in results/frankl-generator-families/:
generate_small_cyclic_weight_orbits.pysmall_cyclic_weight_orbits.json.gzsmall_cyclic_weight_orbits_manifest.jsonreplay_small_cyclic_weight_orbits.py
Two generations with different Python hash seeds produced identical bytes. An independent literal-set replay checked all 27,264 records. Those checks are computational evidence. Lean must still check the finite arithmetic, maps, and complete coverage used by the final theorem.
Checked transport theorem
FranklLocalMaskEquivalence proves exact equality between the mask checker and the local certificate checker under their stated domain conditions. It also proves transport under an injective map of the actual finite local support. No injectivity assumption is made outside that support.
FranklSmallCyclicWeightTransport.transportCheck checks both periods, entry well-formedness, exact map keys, injectivity, the local-generator image, the core image, and the transported weights. Point pairs run from the representative to the target.
transportCheck_sound converts the source's successful window check into the target's successful window check. Its only finite premises are the transport check and the source check. Diagnostic examples reuse the period eight pilot to prove period seven row 23 and row 540. Row 540 uses the explicit replacement vector [0,0,0,1].
The semantic helper passed strict source checks, an explicit import all axiom audit, and exact stock-kernel replay. The gate included 28 modules, 377 project theorems, and the eight reused private pilot chunks. The public helper itself has a 22-module dependency closure. Diagnostic probes are not public proof sources.
FranklSmallCyclicMaskTransport.maskTransportCheck performs the same transport checks with natural-number masks. Its soundness theorem implies the semantic checker above. Three diagnostic maps took 97, 142, and 289 milliseconds of type checking, compared with 352, 346, and 855 milliseconds for the semantic checks. The complete 31-module diagnostic gate audited 386 theorems and replayed the exact closure. This speed comparison does not replace Lean proofs for the full map table.
Checked cached evaluator
FranklLocalMaskCached computes the local generators and inside imbalances once before iterating over boundary pairs. cachedCheck_eq proves equality with the original mask checker for every input.
Its complete local gate checked ten modules and 155 project theorems, then replayed all ten modules with the unchanged pinned stock checker. On the same two difficult period-nine rows, the measured source time fell from 8.80 to 6.48 seconds and stock replay from 10.76 to 9.19 seconds. These are two-row measurements, not a timing guarantee for the full table.
The local-candidate workflow independently rebuilds the complete selected closure, audits every owned theorem, requires each object facet, and checks the exact reported replay set. It records source, object, configuration, gate, and checker hashes. Only propext, Classical.choice, and Quot.sound are permitted axioms.
Hosted pilot evidence
The unchanged literal pilot at commit 77fa9a758cb5753693e66c78f8cde44a09fb4a24 passed the hosted source, all-declaration audit, and stock-kernel gates:
The pilot checks 16 original rows at each of three periods. Word coverage proves a generic overlap enumeration theorem. Neither result supplies the remaining full arithmetic table or the initial graph and rank data.