Skip to content

Autonomous Proof-Carrying Engineering.

How we work, the lab’s current process rather than a result (this engine has run since 10 September 2026; earlier results came before it): AI agents plan and execute the engineering, a separate verifier re-runs a check command named before the work began to decide when a piece of it counts as done, and failures go back for one fix.

See the resultsResults and theorems
  • 2,259Lean 4 theorems and lemmas across the lab’s codebases, not only those behind the results here, that compile with no proof step left as a placeholder, helper lemmas included; the axioms they rely on were not audited
  • 221engineering cycles whose check command, named before the work began, passed when a separate verifier re-ran it
  • 3,452AI-agent sessions that worked in the portfolio's codebases, Dec 2025 – Sep 2026

How the work gets done

3,452 AI-agent sessions worked in the portfolio's codebases between December 2025 and September 2026. Since 10 September 2026 that work has run through an engine in which a separate verifier, not the agent that did the work, re-runs each check command.

  1. Plan

    An AI agent (Claude Opus, in plan mode) writes each unit of work as a plan, and the plan must name — before any work starts — the one check command whose passing counts it as done.

    293plans written

  2. Execute

    The same agent thread executes its plan in the codebase and commits only the files it names.

    305execution runs

  3. Check

    A separate verifier — not the agent — re-runs the check command, named before the work started, with a fixed toolchain and records the full log. The strictest checks rebuild the code in a clean checkout first.

    221cycles whose check command the verifier re-ran and passed, from 473 recorded verdicts

  4. Fix

    A failed check gets exactly one fix pass. If it still fails, the work stops and waits for a human.

    74fix passes

  5. Lock in

    Proofs written in Lean are compiled by the Lean 4 kernel; every other result’s page says how it was checked. Results are counted in stages: from everything the portfolio's codebases have put forward, down to the results whose own check went red when a separate program that did not build the result changed a number the check reads; that program works inside the lab's programme. No third-party organisation has audited any stage. Every stage, and what it rests on, is on the Proof and disclosures page.

    2,259compiled Lean 4 theorems

The limits and disclosures, on the Proof and disclosures page

How we show numbers

Every number on this site links to the file it comes from. How each result is checked

  • We never show a number before its file has loaded.
  • A question we have not checked yet is marked as unchecked.
  • A check that found nothing says so.
  • A file with no value for a question says so.
  • A number whose file is missing or has changed is not shown.
  • Two files that disagree about what a number describes are both flagged.
  • A number from too few samples shows its sample size.
  • Two files that give different values are both shown.
  • A file we cannot publish is listed by its fingerprint only.
  • A measurement more than a week old shows its age.
  • A question that does not apply to a page is left off it.
  • A measurement whose program failed is shown as failed.