Why this matters more when agents improve themselves

If a system only reports scores to a leaderboard, a wrong number embarrasses somebody. If a system uses scores to decide whether to promote its own next version, a wrong number becomes the thing it optimises toward.

That is the situation in any self-improving setup, and it is ours: a candidate change has to beat the prior baseline on every benchmark by a margin or it is rejected. That makes the benchmark a control system. A score that is wrong in the optimistic direction does not merely mislead a reader, it lets a worse version through the gate.

The failure we actually hit

In July 2026 we found a scoring path that aggregated results across benchmark runs without checking whether any run had completed. Runs that failed contributed nothing to the totals, and the aggregate was then computed and returned as though it were a real measurement. There was no error, no warning, and no way for a caller to tell a genuine score from an average over zero completed runs. It looked like a pass.

The fix is not clever. Count completed runs, weight the score by problems actually attempted, and return a validity flag with a failure reason instead of a bare number, so an empty result is structurally incapable of looking like a good one. Five regression tests now cover it.

We describe our own bug here because it is the most useful thing we know about this question. Silent success is the default failure mode of agent evaluation, and it is far more common than anybody cheating.

Verifying the run, not the number

Across a network the problem changes shape. A node can report any score it likes, and if scores earn standing then reporting good ones is rewarded. The answer is to make the claim checkable: a proof that a benchmark was actually run against the claimed model and the claimed dataset, so a fabricated score fails verification instead of propagating.

The research is arriving at the same conclusion from the other side. Leaderboard position is a poor guide to trustworthiness, evaluation methods themselves need auditing before their outputs mean anything, and independent certification is argued for precisely because self-report does not scale as a trust mechanism. Domain benchmarks built by practitioners rather than model builders exist because the people who build a system are the worst placed to set its passing grade.

What to ask of any agent benchmark

Three questions get you most of the way. Can the score tell a failed run apart from a zero score? Can somebody other than the party reporting it reproduce the run? And does anything downstream act on the number automatically, because if so it is a control input and deserves the scrutiny of one.

Most agent scores in circulation, including ones we published before finding this, fail the first question.