Repository version: 3.9.x Purpose: Canonical evidence flow for a new feature or controlled change. Tailor rigor to
L0-L4; do not skip required human or independent roles.
| Activity | Question | Primary skill | Canonical result |
|---|---|---|---|
| Verification | Was the specified output built correctly against the approved baseline? | red-team-verifier |
.agile-v/VERIFICATION_SUMMARY.md, VER-XXXX, test/eval evidence |
| Intended-use validation | Was the right system built for representative users, use, and operational conditions? | validation-agent |
VALIDATION_PLAN.md, VALIDATION_PROTOCOL.md, VALIDATION_REPORT.md |
A verification pass does not establish intended-use validation. Validation is required only when the intended-use and risk context call for it, but it cannot replace failed verification.
| Step | Activity | Evidence and stop condition |
|---|---|---|
| 1 | Classify risk and scope. Record delivery level L0-L4, affected configuration, uncertainty, controls, owner, and residual-risk authority. |
.agile-v/RISK_REGISTER.md; unresolved scope, critical residual risk, or unknown authority blocks progress. |
| 2 | Persist the draft. Convert source intent into atomic requirements with acceptance criteria and typed source lineage. | .agile-v/REQUIREMENTS.md in draft_persisted; chat text is not synthesis input. |
| 3 | Record independent findings. logic-gatekeeper challenges ambiguity, conflicts, constraints, and testability without editing the draft. |
Durable FND-XXXX findings; unresolved blocking findings return to the architect. |
| 4 | Obtain Gate 1 approval. The architect resolves or rejects each finding with rationale; an authorized human reviews the revision and findings. | Approval in .agile-v/APPROVALS.md; rejection returns to revision, not build. |
| 5 | Freeze the baseline. Capture the approved revision and register it. | Immutable baseline plus .agile-v/ARTIFACT_INDEX.yaml; only this baseline is synthesis input. |
| 6 | Build and design tests independently. Build artifacts implement the baseline; Test Designer derives tests from the baseline without reading implementation. Use separate fresh contexts, or run test design first if one context is unavoidable. | ART-XXXX -> implements -> REQ-XXXX@revision; TC-XXXX -> verifies -> REQ-XXXX@revision. |
| 7 | Verify independently. Red Team executes tests, challenges artifacts, records VER-XXXX findings, and requires re-verification after fixes. Requirement changes go through a change request and a new Gate 1 baseline. |
.agile-v/VERIFICATION_SUMMARY.md; unresolved verification failures block Gate 2. |
| 8 | Run the Eval Gate. Execute required offline/online suites against policy thresholds. | .agile-v/EVAL_RESULTS.md; status must be PASS or an authorized WAIVED, and the EvalGate block belongs in VERIFICATION_SUMMARY.md. |
| 9 | Perform intended-use validation when required. Use representative users, environments, configurations, and data; reconcile anomalies and residual risk. | VALIDATION_REPORT.md; insufficiently representative evidence cannot support intended-use acceptance. |
| 10 | Complete AI influence evidence. For material AI influence, record model/runtime identity, loaded skills, tools, context sources, affected artifacts, hashes, tests, and evidence locators without secrets or hidden reasoning. | .agile-v/aibom/<task-id>/AI_RUN_MANIFEST.yaml linked to the evidence bundle; apply approved policy before using the preview agile-v-aibom skill. |
| 11 | Obtain Gate 2 and release. Present verification, applicable validation, Eval Gate, control conformance, trace coverage, open anomalies, rollback, and residual-risk evidence to the authorized human. | Gate 2 approval is durable; release-manager then executes the approved rollout/rollback plan. No autonomous production release. |
The normative state transitions and typed edges are defined in Canonical Lifecycle Contract; minimum rigor by level is defined in Risk Classification.
Each step above has a concrete, schema-validated example fixture in examples/golden-journey/ (tests/test_golden_journey.py re-validates all of them). examples/negative/ demonstrates the corresponding rejection paths (stale evidence, self-approval, expired approval, policy drift, verifier contamination, an Evolve goalpost change, and others), each stating plainly whether it is backed by an executable test or is a specified-but-not-yet-tested rule.