PROBLEM
Generated or learned behavior can look plausible while still requiring a named human authority before it is allowed to affect users.
USE CASE
Human approval for AI behavior changes
Keep approval as a distinct authority step instead of treating generated candidates or passing evaluations as automatic permission to release.
OLD WORKFLOW
The same workflow that creates a candidate often implicitly decides that the candidate can ship.
EXAMPLE
A learned support escalation rule may pass evaluation but still require an operations owner to approve the release scope before canary traffic starts.
VIRA WORKFLOW
Keep evidence, authority, evaluation, and release separate.
- 01
Create evidence and candidate behavior without activating it.
- 02
Run exact evaluation against the baseline.
- 03
Route the governed decision to an authorized reviewer where required.
- 04
Keep approval identity separate from runtime health.
- 05
Promote only after all required authorities are satisfied.
WHAT IS GOVERNED
- Approval capability
- Candidate state
- Release authority
- Audit trail
WHAT IS MEASURED
- Reviewer decision
- Evaluation state
- Promotion evidence
- Deployment events
WHAT CAN FAIL
Failure states are part of the design.
- Approval is missing or stale.
- The approved artifact no longer matches the candidate identity.
- A later source change invalidates the release evidence.