PR Evidence Pack Showcase

Evidence packs make AI-assisted delivery reviewable.

AI-assisted delivery changes the review problem. A payment-platform PR can look complete before anyone has proved the right risks were checked. The evidence pack ties each release risk to the proof required for governed delivery review.

Discuss design partnership Review evidence model

Value

From review confidence to release evidence.

Pain	Without evidence pack	With evidence pack
Risk review	Reviewers infer risk from comments and code shape.	Risk, controls and evidence are explicit.
Security questions	Proof is requested late.	Required evidence is attached before approval.
Release ownership	Approval becomes a binary ship/no-ship call.	Decision, residual risk and rollback triggers are visible.
Incident review	Evidence is reconstructed after the fact.	The decision record exists at merge time.
Agent output	Confident summaries can hide missing controls.	Agent findings become structured review signals.

System model

One repeatable record for review, release, audit and aggregation.

The artifact is the output. The product value comes from producing that output consistently across high-risk PRs without hand-building a release packet each time.

PR diff Risk classifier Required evidence Review signals Release decision Audit record

Delegation boundaries

Which AI-assisted changes are safe to delegate is tracked by risk class.

Review bottlenecks

Missing-evidence categories are visible across PRs.

Control gaps

Repeatedly absent controls are aggregated, not rediscovered.

Incident readiness

Review starts from the release evidence record.

Proof status

What this showcase demonstrates.

Claim	Status	Next validation
Evidence categories are useful	Demonstrated by representative pack	Run against multiple high-risk PRs
Release judgment can be expressed	Demonstrated by payments example	Validate against review threads and CI output
Packs can be generated repeatedly	Product hypothesis	Wire generator to PR metadata, CI and review comments
Aggregation creates governance value	Product hypothesis	Measure missing controls, exceptions and review gaps

Change record

Webhook signature rotation for payout events.

High integrity · medium availability · staged rollout · sanitized, representative.

added:
  - key_id_header_validation
  - dual_key_verification_window
  - timestamp_tolerance_check
  - replay_nonce_store
  - audit_event_for_rejected_webhooks
unchanged:
  - payout_execution_logic
  - ledger_write_contract
  - operations_status_schema

Incoming webhook Known key id? Timestamp valid? Nonce unused? Signature valid? Accept or reject + audit

Evidence matrix

Risk mapped to proof.

Risk	Control	Evidence	Status
Forged webhook marks payout complete	Signature and key id verification	Negative signature tests	Controlled
Replay changes payout status twice	Nonce store and idempotent worker	Replay test matrix	Controlled
Clock skew rejects valid events	Timestamp tolerance and alert	Manual release review	Accepted
Logs expose sensitive payload	Redacted audit schema	Static review	Controlled

Pack object

The release decision as structured data.

The pack is concrete enough for PR review and structured enough for later aggregation.

{
  "id": "payout-webhook-signature-rotation-v1",
  "risk_class": "high_integrity_medium_availability",
  "validation": {
    "unit": "pass",
    "integration": "pass",
    "contract": "pass",
    "manual_release_review": "pass"
  },
  "release": {
    "decision": "approve_staged_rollout",
    "rollback_available": true
  }
}