Blog

Field notes with the receipts attached.

What OKReport AI built, what we checked, and what we would not claim without reproducible evidence. The client stays private; the result does not.

Subscribe by RSS

How we publish

Three rules decide whether something on this blog is a claim or a story.

  1. Verified, not summarised

    A result enters a benchmark only after a reviewer checks the available evidence. What could not be reproduced stays visibly unreproduced.

  2. Sanitized before publication

    Client identity, internal addresses, customer records, private source, and credentials are removed. Failures and human help are not.

  3. Concise by default

    Each benchmark starts with the useful conclusion, then lets readers expand the implementation details, prompt, and evidence.

Where this series goes

Not release dates—the questions the series is set up to answer next.

  • Same model, different harness

    Hold the model constant and change the agent harness to separate model behaviour from the tooling around it.

  • A brief with a migration

    Add schema change and rollback to a real feature, where skipped verification becomes visible quickly.

  • Frontend-under-pressure runs

    Keep publishing authentic interface evidence alongside the code and test outcome.