An experiment in AI-driven theory generation
The question is whether an AI-driven research loop can produce mathematical physics that survives machine checking and blind testing. It has run once. It is running again, differently.
Run #1 — archived
Roughly 388 sessions, claiming 63+ physical constants with zero free parameters. Audited in September 2026: its two flagship results sit 1,755σ and 3,261σ from CODATA 2022, and those fractions are continued-fraction convergents of the measured values — which is what fitting produces, not what a theory predicts.
It is kept intact and readable rather than deleted. A negative result with its records intact is worth more than a quiet retraction, and the record of how it went wrong is the input to run #2.
Read the archived site →Run #2 — in progress, not yet public
Restarted from close to scratch. Lean 4 proofs with verdicts pinned to source hashes instead of scripts that could not fail; a sealed, multi-provider decoy experiment instead of predictions compared against values already known; rules enforced as gates with their own self-tests instead of written as prose; derivations structurally barred from reading the measurement register.
It becomes public only after a scrub gate passes and a deliberate release decision — not when it looks ready.
Nothing from run #1 carries over as a claim. Its formulas do not transfer; only its epistemic record does.