Change Control

Critique is an AI Change Control Platform. Instead of only leaving review comments, it sits at the merge boundary and records a durable system of record for every pull request: where the change came from, how risky it is, what evidence backs a blocking decision, whether policy allows the merge, whether a repair was verified, and what past incidents taught the rules.

Reviews still run. When policy allows, Critique executes AI-powered sandbox reviews (and other review paths) and stores the result as an evidence run linked to the passport. Change control does not remove review — it subordinates review to control: who gets a full pass, why it runs, what must not merge, and what happened last time.

Deep dive: Change Control launch essay. Marketing gate overview: critique.sh/checkpoint. Operator surfaces: Dashboard → Passports and Dashboard → Control Board.

For production user reports before a PR exists, use Critique Intake to turn a complaint into an investigated packet. Once that work becomes a PR, the Change Passport records provenance, review evidence, merge policy, and proof.

The control loop

flowchart LR
  Agent[Agent or human opens PR] --> Passport[Change Passport opens]
  Passport --> Risk[Provenance + risk score]
  Risk --> Gate[Agent Firewall gate]
  Gate --> Review[Evidence-backed review]
  Review --> Policy[Merge policy: dry-run to enforce]
  Policy --> Repair[Verified repair if needed]
  Repair --> Merge[Merge]
  Merge --> Memory[Incident learnings feed back into rules]
  Memory --> Risk

Review-First vs Change-Control Orientation

Change Passports

A Change Passport is the PR-level record. The queue at /dashboard/passports lists every governed pull request with repo, author, a source badge, gate outcome, risk, verdict, evidence status, merge policy, proof, and memory. Filter by repo, risk, state, and verdict.

The detail view at /dashboard/[owner]/[repo]/pulls/[number] renders the full passport:

When no immutable snapshot exists yet, the passport infers provenance live from PR signals and labels it as heuristic so the record is never blank while the first review runs.

Compliance export: Use Export audit bundle on the passport detail header, a row on /dashboard/passports, or Export filtered after applying queue filters. Critique returns signed, redacted JSON (manifest hashes, snapshots, policy, incidents, timeline)—formatted for SOC 2 / ISO evidence handoff, not a certification. API keys with read:passports can call GET /api/v1/passports/:passportId/export; signed-in automation can batch up to 25 ids via POST /api/passports/export. Programmatic review queueing and verdict polling: Merge Gate API.

Agent risk scoring

Every review run persists a risk score, a band (low, medium, high, critical), and human-readable reasons. Risk flows into the passport queue, the passport detail, and the control-room overview, so it is a first-class column you can filter and sort, not a sentence buried in review prose.

Evidence Contract

The Evidence Contract normalizes review artifacts into a consistent shape and exposes:

• the blocking decision for the run, and
• the evidence linked to each finding.

The rule is simple: if a change is blocked, the block must point at something. Legacy review artifacts are normalized by the same accessors, so older runs still render under the contract. Drill into the evidence from any review-run page.

Evidence runs (AI-powered sandbox review)

An evidence run is a commit-level review linked from the passport. Critique still performs deep review when gate and review policy allow — typically via managed sandbox execution, model routing, scout plus specialist lanes, and a lead verdict. The UI label changed from “the product” to “evidence on the passport,” but the engineering work of reviewing code did not stop.

Use Dashboard → Review runs for commit-level drill-down; use Passports for PR-level control.

Merge policy as code

Merge policy is a schema, an evaluator, and the published Critique / Merge Policy check. Configure it in the dashboard (Control Board → Change policy) or in a repo file.

Modes

Repo-file policy

Critique reads policy from .critique/policy.yml, .critique/policy.yaml, or .critique/policy.json when present. Choose repo_file as the source on the Control Board to let the repository own its policy.

Natural-language policy editor (v5+)

On Control Board → Change policy, describe rules in plain English. Critique compiles them into strict policy JSON with an LLM assist (preview, YAML, rule diff, confidence badge), then evaluates merges with the same deterministic engine — no model calls at merge time.

Supported v2 rule shapes include path-aware blocks/warns, required test-file touches, minimum current-head GitHub approvals, and safety modes (Require review, Allow draft PR, Ask followups) before apply or policy PR.

Save paths: persist to dashboard, open a branch that writes .critique/policy.yml and opens a PR, or copy YAML manually.

Operator overrides

An operator can override a policy decision from the passport. The override records provenance (who, when, and why) and can patch the GitHub check-run status so branch protection reflects the call you made. Override errors (for example, a failed check-run update) are surfaced on the passport.

Verified repair (Remedy proof)

When Remedy fixes a finding, Critique stores a proof bundle on the attempt:

• a patch hash,
• validation results,
• the push or export mode, and
• a verification linkage back to the review run that confirmed the fix.

A repair is recorded as done because the proof says so, not because an agent claims it is.

Findings memory and incident learnings

• Findings memory surfaces suppressions and a feedback ledger on the Control Board, with revoke and expire actions.
• Finding feedback on review runs — From the dashboard, mark findings Accepted, False positive, Fixed, or Suppress without leaving the run. Feedback feeds memory and the Skill Performance Leaderboard.
• Model-feedback opt-in (strictly optional) — Grant or revoke consent per installation (or single repo) under Control Board → Memory. When active, you can share anonymized examples (patches and secrets stripped) via finding actions or POST /api/model-feedback/exports. GitHub slash commands accept --share-feedback only with the same consent gate — no silent export from PR comments.
• Incident feedback ingests events from Sentry, Linear, Jira, Vercel, and generic or manual sources, links them to the passport, and drafts learnings you can promote into control rules or dismiss.

This closes the loop: what breaks in production teaches the gate what to catch next time.

The Control Board

The Control Board at /dashboard/control is one operator surface with five tabs:

Legacy /dashboard/change-gate and /dashboard/checkpoint redirect into the Control Board gate. The legacy automation editor is preserved directly for anyone who still needs it.

Relationship to the rest of the platform