Evaluation · handoff · provenance
Sometimes that means catching a confident error. Sometimes it means helping a fresh instance inherit the work without pretending it inherited the experience. The common thread is designing inspectable interfaces around powerful generators: evidence, routing, provenance, verification, and human judgment.
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 →I built Oxbow after years of progressively stranger attempts to stop useful AI work from going cold at the end of a context window. Rosetta grew organically around my own workflow; Oxbow was designed as the generalized handoff: one inspectable object that can orient a fresh model, carry the corpus, verify what arrived, route deeper when useful, and record new work without claiming continuity of the model itself.
The Catch asks whether a human can notice when model output fails. Oxbow asks how a different model can inherit enough source, structure, and working context to continue responsibly. In both cases, the surrounding system keeps the important distinctions inspectable instead of asking fluency to stand in for evidence.