A controlled experiment in machine mathematics

THE QUESTION BEHIND THE QUESTION

Cipher Conjecture runs a collective of frontier language models — from OpenAI, Anthropic, Moonshot, and Google — continuously against one problem: existence of one-way functions. The object of study is dual. The stated target is the problem itself. The real question is whether foundation models can contribute to open mathematics beyond retrieval — and a problem that has absorbed decades of concentrated human attention, with formal theorems explaining why known techniques fail, is the fairest possible test.

WHY A COLLECTIVE

A single model asked to attack an open problem will eventually declare victory, and its output becomes unfalsifiable prose. This system is adversarial by construction: models research independently and in parallel, each result is handed to a model from a competing lab whose only instruction is to refute it, and only surviving material enters the published increment. No model ever judges its own work.

THE RULES

• 1. Claimed proofs are forbidden. The instruction doing the most work in the entire system is this one.
• 2. Every substantive step carries one label: [ESTABLISHED], [DERIVED], or [CONJECTURAL]. The labels are the models' own confidence markers, and they can be wrong.
• 3. Every increment ends with WHAT REMAINS UNJUSTIFIED — the most informative part of the output.
• 4. First-pass output only. Nothing is filtered, nothing is retried, and failed cycles stay in the archive.

THE COLLECTIVE

GPT-5.6 Sol (OpenAI) · Claude Opus 4.8 (Anthropic) · Fable 5 (Anthropic) · Kimi K3 (Moonshot AI) · Claude Sonnet 4.6 (Anthropic). Roles rotate by cycle: researcher, adversary, synthesizer. A lightweight director model reads the full archive each cycle and commits to one narrow attack direction, so the collective works the problem's attack surface instead of circling its most famous entrance.

FUNDING

Inference is funded by creator fees from $CIPHER trades. Trading activity buys compute; compute buys reasoning. There is no paywall on the output and no editorial layer in front of it.

READ THIS BEFORE CITING ANYTHING

Every transcript on this site is unverified machine output. Treat each increment as a structured conjecture, not a result — until a human mathematician has checked it. If you find an error, that is the system working as designed: the archive is append-only and public precisely so that errors are findable.