Acceptance criteria
Acceptance criteria are the quality bar a trained model must clear before it can reach production. You define them once, on the capability; the platform enforces them every time, for every candidate model, forever — including models produced by automatic retraining and fine-tuning.
This is the single most important governance control in Pradra Studio: no model ships without passing the gate, and no path around the gate exists.
How a rule works
Each criterion is a simple rule: a metric, a comparison, and a threshold.
| Part | Example |
|---|---|
| Metric | accuracy, f1, precision, recall, rmse |
| Operator | >=, <=, >, <, == |
| Threshold | 0.85 |
A capability can carry several rules — "accuracy at least 0.85 and recall at least 0.75". Every rule must pass. If even one fails, promotion is blocked and the Studio tells you exactly which rule failed and by how much — even if the model improved on the headline metric.
For generative capabilities the same mechanism reads generative metrics (golden-set scores, structural validity, safety rates) — see Generative evaluation. For agents it reads task success and violation counts — see Agents. One gate, many kinds of intelligence.
Where the gate applies
- Publishing a model to the registry.
- Promoting a model to be the capability's production version.
- Approving a retraining candidate — the check is re-run at approval time, not trusted from earlier.
- Activating a generative prompt or fine-tuned model.
A candidate that fails is marked rejected — which is worth reading carefully: rejected means the controls worked and production was untouched. It is not an error; the previous model keeps serving.
Choosing good criteria
- Anchor on the business. "Accuracy ≥ 0.85" is meaningful when you know what an error costs. Start from the cost of a wrong prediction, not from what the model happens to score today.
- Guard the failure mode you fear. For fraud detection, a recall floor matters more than overall accuracy; for a spam filter, precision protects your users' inboxes.
- Set the bar before you train. Criteria chosen after seeing the score aren't a gate — they're a rubber stamp.
- Empty criteria pass everything. That is allowed (useful while prototyping), but a capability heading to production should always carry at least one rule.
Setting criteria in the Studio
- Open the capability's detail page.
- Edit its acceptance criteria: add one row per rule (metric, operator, threshold).
- From then on, every evaluation report is checked against these rules. The result is visible on the model's evaluation page — pass or fail, rule by rule.
Next
- See where the scores come from: Models & evaluation.
- See the gate at work in automation: Drift & retraining.