Function, integration, performance, deployment, rollback
Readiness as policy
Replace the generic checklist with explicit release logic.
Versioned policy selects obligations from release facts, tests evidence and thresholds, records exceptions, and explains the recommendation.
2 blocking actions remain
Multidimensional readiness
A healthy average must never hide a critical failure.
Dimensions remain separate, and blocking controls cannot be averaged away. Policy determines which evidence is required for this specific release.
Support approval
Training threshold
Task success, tool use, grounding, safety, escalation
Identity, access, privacy, retention, audit
Monitoring, support, incident handling, continuity
Role impact, communication, training, feedback
Baseline, target, source, period, benefit owner
Policy-as-code
Release facts become obligations—consistently and audibly.
A visual editor can make policy accessible, while the underlying representation remains versioned, testable, and reconstructable.
when: agent.has_write_access == true
require:
- security_review == approved
- rollback_test == passed
- human_escalation == validated
when: affected_users >= 500
require:
- support_plan == accepted
- training_plan == acceptedFrom fact to decision
Five distinct records preserve accountability.
The system recommendation and the human decision are deliberately separate.
Every exception names the unmet rule, accepting authority, reason, compensating action, scope, and expiration.
Start with one real release