ModelCensusopen-source ai reliability harness
Blog20 Aug 2026methodharnesspost-mortem

We accepted 2,900 failures without reading them — and said so in the data

A gate everyone routes around isn't a gate. So we named the shortcut instead of pretending.

Publishing a panel run means clearing a gate: every failing probe reviewed by a person. On a sixteen-model run that is 2,902 probes.

Nobody reads 2,902 transcripts. We did not.

Expectation versus reality

The gate was designed assuming review would be per-probe and careful. That holds for a five-model pilot. It does not survive contact with a twenty-model panel.

What happens to an unrealistic gate is not that people stop publishing. It is that they satisfy it nominally and the record cannot tell.

The unqualified flagMarks failures reviewedClaims inspection happenedRecord can't distinguishvs.Bulk acceptMarks failures reviewedStamps every probereviewed_by: bulk-acceptRecord says so plainly

Diagnosis

The gate had one setting and reality had two

Careful per-probe review and blanket acceptance are both legitimate decisions in different situations. The system only modelled the first.

Why: When a system models one behaviour and people need two, the second one happens anyway and goes unrecorded.

We nearly built decision support that decided

An early triage tool was one commit away from auto-approving 'obvious' cases.

Why: That would have been the bypass the invariant exists to prevent, wearing a helpful name. It now re-checks each failure against the same code path the detector used and approves nothing.

What we do now

Two paths. The careful one, and a bulk path that writes its own name into every probe it touches. Anyone auditing a published loop can query which failures were individually inspected.

The last census: 2,902 probes, bulk-accepted, stamped. It is in the data and it is in this post.

The honest version of a shortcut is a shortcut that leaves a trace.

Which of your quality gates does everyone satisfy nominally — and would you know?

Every figure here describes something measured and committed. See the measurements · read the method