LOADING
Core explanatory content is server-rendered. This state is only for a route segment that is still resolving.
EVALUATIONS
Compare an exact candidate with an exact baseline, surface critical regressions, preserve provider execution evidence, and seal what a release decision actually evaluated.
DEFINITION
Evaluation is evidence about a specific candidate and baseline. Passing evaluation can make a revision eligible for the next release step, but it does not automatically grant production authority.
HOW IT WORKS
Bind the evaluation to an exact candidate behavior revision and exact baseline.
Select the explicit dataset, visible cases, risk partitions, and holdout evidence required for the decision.
Execute the candidate and baseline against the intended provider or model target.
Separate aggregate quality from critical regressions that can block release even when an average score improves.
Record deterministic checks, model-assisted judges, and human review according to the risk of the behavior.
Seal the evaluation evidence so a later release can prove it is promoting the same candidate that was evaluated.
PRODUCT PROOF
These fields describe the product evidence Vira preserves for this part of the lifecycle.
CONTROL BOUNDARIES
Candidate, baseline, case set, target, and relevant execution identity are named rather than inferred.
Hard-policy weakening, forbidden actions, broken verification, or missing escalation can block release independently of aggregate scores.
Model-assisted scoring is evidence, not unquestionable authority; deterministic and human evidence can remain mandatory.
The release path can verify that the revision and artifact being promoted match the evaluation that justified eligibility.
LIMITATIONS & FAILURE STATES
DEVELOPER HOOK
Keep provider, behavior revision, and evaluation inputs explicit so the deployment stage can verify that promotion evidence still matches the candidate.