Feedback and realignment

Corrections are how your model gets better. Flag an output, say what you would have said, and the next alignment cycle incorporates it, then re-attests against your values.

Submitting corrections

On a model's dashboard page, rate any output up or down. For a thumbs down, add the response your business would actually give: “we'd never say that, instead we'd say this.” Each correction lands as pending feedback.

Realignment

A realignment consumes every pending correction as new control signal and starts a fresh cycle. The model walks through the same stages as the first alignment, including policy attestation, and your corrections are marked incorporated.

ready ──▶ corrections pile up (pending)
                    │
            realignment called
                    │
queued ──▶ training ──▶ evaluating ──▶ ready
                    │
        corrections marked incorporated

Corrections accumulate, a realignment consumes them.

Nothing changes for callers

The endpoint URL and API key survive realignment. The new policy artifact is promoted behind the same address once evals pass, so integrations keep working without a deploy on your side.

Batch corrections and retrain when a real pattern has built up. Retrains are billed as training tokens, see usage and billing.