Three brains. One chain.
The loop, running.

Fair but hard: certified unique-solution puzzles generated by language models

The generate → encode → certify loop: LLMs write puzzles; two independent encodings are proved equivalent with z3 (dual encoding); a SAT solver proves exactly one solution; a five-level deduction ladder grades every item with minimal-core explanations. A reasoning benchmark whose items carry solver certificates — fresh-generated, contamination-proof, machine-certified.

The evolution we build together: her certification pipeline meets our receipts — every certified puzzle becomes a khipu receipt; her benchmark becomes a third proving-ground axis (models generate, z3/SAT certifies, the certificate chains). Certified-benchmark-evaluation, receipted end to end. The repository →

The scoring protocol & the figure/data contract

The wet-lab collaboration: a declared mapping contract between figures and their underlying data — verified mapping, with the honesty rule that MATCH does not establish measurement authenticity. Methods and fixture schemas cross the collaboration boundary; recordings, unpublished values, student work, and strain designations do not.

Instilled in the estate: the protocol is now doctrine — frozen acceptance criteria before data inspection, hashed machine output before human review, independent verification that never shares the suspect logic, fail-closed on empty denominators. The brain trains on it; the proving ground applies it.

Quality-vs-precision curves admitted only from verifiable run receipts. Fixtures, projections, and BLOCKED genesis records are never benchmark rows. Unverifiable runs appear as absent — fail-closed display. Promotion compares only the same weights-revision and quality metric.

The guarantee: every admitted point is hash-chained into the run ledger — editing or reordering a receipt breaks verification.

Every advisory signal ships with a DSSE-signed receipt a third party can verify offline.

The honest scope box: advisory research, paper only, not financial advice. No order execution, no custody — those code paths do not exist. Backtests are measured replays of real history and predict nothing. Confidence capped at the 0.97 trust ceiling — nothing reaches certainty, by law. Λ uniqueness = Conjecture 1, open, stated as open.