Performance baseline — the locked number

The short answer. A performance baseline is the number a model was approved at — the AUROC or equivalent, measured on local data, recorded with its date and owner, and then locked. Locking is the point: monitoring means comparing today against the approval-day number, and a baseline that can be quietly revised makes every later 'still performing' claim unfalsifiable.

Every monitoring conversation is secretly a comparison: against what? The baseline is the answer — the performance measured at local validation, on your population, recorded at approval.

The governance property that matters is immutability. If the baseline can be edited after the fact, “the model is within threshold” is a claim nobody can falsify, because the goalposts have a handle on them. Serious record-keeping therefore locks the baseline: supersede it with a new dated entry if revalidation justifies one, but never revise it in place. That is literally how HAI-OS stores baselines — insert-only, with supersession — so the question “what was this model approved at, and when?” always has exactly one auditable answer.

Where do you actually stand? The free RUAIH readiness score maps your organisation against the five focus areas in about eight minutes, and the published crosswalk shows how each control lands across RUAIH, CHAI and the NIST AI RMF.

Performance baseline — the locked number

Keep reading