Research mathematics needs proof assurance, not a single judge score

What the laboratory reports

• OpenAI reports attempts on research-level mathematical problems and distinguishes the sprint from a controlled evaluation.
• One attempt initially viewed as likely correct was later judged incorrect after expert and community analysis.
• The report calls for a more rigorous framework for evaluating research-level mathematical work.

The mathematical problem

A proof is a dependency graph, not a scalar score. Its reliability depends on semantic equivalence, hypothesis coverage and the joint false-acceptance rate across necessary steps. Correlated generator and verifier errors invalidate naive independence assumptions.

GERO's proposed response

• Compile the argument into assumptions, imported results, lemmas, calculations and conclusion dependencies.
• Route nodes to heterogeneous checkers and reserve expert review for semantic bridges.
• Publish bounded statuses and replayable evidence rather than an unqualified correct/incorrect label.

A falsifiable experiment

• Build a sealed set of valid and subtly corrupted research-style proof trajectories.
• Compare same-family judging, cross-family voting, formal tools and the GERO graph protocol.
• Use false acceptance of invalid proofs as the primary risk metric.

Read the primary source

This article is an original analytical summary, not a republication. Read Our First Proof submissions for the laboratory's complete claims, methods and context.