Sulayman Bowles

Make AI results inspectable

Agent evaluation, workflow design and exact search, with explicit tasks, replayable records and independent verification.

For agent evaluation, retain the task, tool actions, observations and failure conditions. For an exact search, distinguish a candidate from an independently verified solution.

The first AI managers follows decisions in shops, cafés and radio stations. The ViralBench proposal and replayable-trace study ask how to test agent work; the exact-search case shows what changes when the answer has a precise, independent verification condition.

Reading path updated .

  1. The shopkeeper in the machine

    Examine how people assign, supervise and evaluate AI work inside a workflow.

    15 min read

  2. A proposed Codex harness for ViralBench

    A proposed Codex improvement loop for ViralBench, using retained traces, replay and independent evaluation before release.

    29 min read

  3. Checking Jane Street’s official route

    Follow an exact search from constraints and candidates to a separately checked answer.

    11 min read

  4. Every ordering, from zeros and ones

    Follow a Codex-assisted mathematical construction from an explicit request to a full proof, with six exact rational certificates and independent automated checks.

    5 min read

  5. Two recurrence proofs: a matrix inverse and a binomial transform

    Inspect two written recurrence proofs, exact independent programs and the boundary between automated checks and human review.

    6 min read

  6. Divisibility patterns and the size of a witness

    Compare a sharp family theorem with a universal bound and see precisely which growth questions remain open.

    8 min read

  7. A double-binomial recurrence and its continued fraction

    Inspect the formal-tail proof and exact recurrence checks in a separate note.

    3 min read

Browse all writing