Evaluation · falsification · provenance

I build the layer that catches when AI is confidently wrong.

The common thread is putting inspectable constraints around powerful generators: nulls, provenance gates, confidence boundaries, parity checks, reversible transforms, and explicit falsifiers.

77-item instrument

The Catch

A human-in-the-loop judgment instrument that separates error from suspicion, handles partly-right cases, tracks provenance and solvability, and scores sensitivity and response bias rather than rewarding raw flag counts. The 77-item reference bank is intentionally private.

Read the method →
Governed evidence system

OXBOW

A governed bundle, append-only witness, falsification layer, River Gauge, and codec laboratory where lawful form is not confused with truth and unsupported patterns are allowed to die.

Read the architecture →
Design rule

The generator does not get to grade its own homework.

Across these systems, proposals can be fluent, useful, or surprising without becoming authoritative. Promotion requires something outside the generator to agree: a source document, a null, a baseline, a parity test, a reversible transform, or an explicit human judgment that itself stays inspectable.