Evals 2.0 — now with counterfactual traces

Models that explain what they just decided.

Cortex records every inference with its inputs, its retrieved context and its confidence, so when a decision is challenged you have an answer rather than a shrug.

Drift, before it costs you

Distribution shift flagged against a rolling baseline, per feature, not per model.

Counterfactual traces

Change one input, see what the output would have been. No retraining, no guessing.

Audit-ready by default

Immutable log, seven-year retention, exportable in the format your regulator asks for.