Implementation evidence¶
P11 checks whether completed implementation still matches one exact P10 approval. It reads the approval and evidence request, compares the approved baseline with local Git state, and hashes only the evidence files named in the request.
It does not implement changes, run tests, read application source, write files, commit, push, deploy, call a network, or authorize a release action.
Prepare the request¶
List every current changed file with its exact output from:
Every approved verification command must appear once and in approval order.
Use not-run with empty timestamps and null exit code when a command was not
run. Accessibility, responsive, and browser observations must always appear;
use unverified rather than omitting missing proof.
{
"kind": "design-ai-implementation-evidence-request",
"schemaVersion": 1,
"consumer": "implementation-agent",
"implementationStartedAt": "2026-07-15T12:01:00.000Z",
"implementationCompletedAt": "2026-07-15T12:20:00.000Z",
"executedWork": [
{
"statusEntry": " M src/settings/view.tsx",
"path": "src/settings/view.tsx",
"summary": "Clarified the approved settings action without changing its data flow."
},
{
"statusEntry": "?? evidence/test.log",
"path": "evidence/test.log",
"summary": "Recorded the approved test result."
}
],
"verificationResults": [
{
"command": "npm test",
"status": "pass",
"startedAt": "2026-07-15T12:10:00.000Z",
"completedAt": "2026-07-15T12:11:00.000Z",
"exitCode": 0,
"summary": "Unit tests passed.",
"artifacts": ["evidence/test.log"]
}
],
"observations": [
{
"id": "accessibility-check",
"category": "accessibility",
"status": "unverified",
"summary": "Accessibility was not exercised in this run.",
"artifacts": []
},
{
"id": "responsive-check",
"category": "responsive",
"status": "unverified",
"summary": "Responsive behavior was not exercised in this run.",
"artifacts": []
},
{
"id": "browser-check",
"category": "browser",
"status": "unverified",
"summary": "Browser behavior was not exercised in this run.",
"artifacts": []
}
],
"remainingRisks": [
{
"severity": "p2",
"summary": "Responsive and browser behavior still need confirmation."
}
]
}
Evidence artifact paths must be exact relative paths covered by the approval's
files.change or files.generated selectors. Each file must be regular,
non-symlinked, and remain inside the approved repository root.
Run the review¶
design-ai review-evidence implementation-scope-approval.json \
--request implementation-evidence-request.json \
--target-root /absolute/path/to/repo \
--consumer implementation-agent \
--json > /tmp/design-ai-implementation-evidence.json
The shell redirection writes the output; the operation itself writes nothing.
The result is:
evidence-completewhen scope, Git state, verification, observations, and risks contain no gap;attention-requiredwhen verification failed or was not run, an observation is unverified, pre-existing work overlaps, or a risk remains;blockedwhen baseline identity drifted, a changed file is missing or outside scope, an artifact is unsafe or unavailable, or the record contradicts Git.
MCP and Website Console¶
MCP exposes design_ai_review_evidence. It accepts absolute approvalPath,
requestPath, and targetRoot values plus the matching consumer. The tool uses
the same in-process operation and does not spawn the CLI.
Website Console imports the resulting JSON, revalidates its exact nested approval and request sources, preserves original bytes for export, and restores the approval when evidence is cleared:
implementation evidence -> scope approval -> scope proposal -> target intake
Release boundary¶
P11 records evidence before commit. It does not grant commit, push, deployment, migration execution, or external-write authority. A changed HEAD expires the P10 approval and blocks P11; create a new intake, proposal, and approval instead of editing the evidence around the drift.