Choose an output-check recipe

Bring any verifier logic. Keep rules, models, compliance checks, proof engines, or completed workflow results in your own service. Millwork sends candidate output to its endpoint and records the verdict, quality score, and optional named checks with the run. Choose a recipe by the decision you need; each recipe names what its evidence can prove. If the agent can write the evidence, choose records or a signed grade that it cannot alter before treating the result as independent proof. Already have a compatible evaluator? Test its endpoint locally, then connect it and inspect the first receipt. The Millwork CLI also includes output-check examples and local tests. Run a generated example before changing it; no organization key or endpoint is needed for that local test.

Choose a recipe

Start with	Included example and next step
An existing HTTPS check	Connect the endpoint you have. Skip the generated sample.
Fields, types, or exact math	Check structured output. Includes JavaScript invoice rules and local cases; replace the rules with yours.
Labels, scoring, routing, or citations	Add a semantic judgment. Includes an offline citation example; connect your evaluator later.
A citation-support decision you want Jev to judge	Check citations with Jev. Start with the offline sample, then connect TypeSafe’s model.
A DSPy program whose answers need a Jev check	Run a DSPy program with a Jev check. Python only: an agent endpoint and a check endpoint in your app, offline cases first.
Ranking, moderation, or an existing policy evaluator	Connect an existing evaluator. Includes an offline adapter; connect your service later.
Artifacts and tests from an agent or pipeline	Verify agent or pipeline completion. Includes a local evidence sample; connect your trusted records later.
Long-running research or grading	Dock a Temporal quality gate. Finish work upstream; the output check reads signed evidence and returns a bounded result.
A completed CodeRabbit review of a pinned commit	Check a CodeRabbit review. Adapt Recipe D with offline cases and a trusted review record.
A supplier quote another agent will act on	Verify an agent-to-agent quote handoff. An offline example with synthetic mail; the receiving agent verifies a signed record before acting.

To design the authority, record and receiver checks for another business workflow, follow Build a verifiable agent handoff. The Temporal example is one worked form of the reusable dock long-running work pattern. The check does not wait for a workflow. Each included CLI recipe selects one example from the same kit. The Jev page adapts the semantic-judgment example; the CodeRabbit page adapts Recipe D with --recipe d. Neither adds a new CLI recipe value or requires Temporal. Use Connect an output check afterward for deployment, access, connection, and receipts.

Try an included example

With Node.js 20 or newer, install the supported CLI, then check its version:

millwork --version --json

Use the pinned version from the install guide. The cookbook examples need version 0.1.18 or later. A recipe’s first command copies an existing check into your directory; its local test shows a pass, a rejection, and an unavailable case. Those tests do not approve or start a paid Millwork run. Verdict rule: is_correct is the hard decision. quality_score grades quality from 0 to 1 and cannot override rejection. Timeout, unreachable endpoint, authentication refusal, or malformed response means no verdict.

Example: a successful check response

{
  "is_correct": true,
  "quality_score": 0.84,
  "anchor_results": {
    "required_fields_present": true
  }
}

anchor_results is optional. Its named booleans explain which exact rules passed or failed. Read the endpoint contract before implementing a check.

Who does what

Role	Owns
You	Rules, private key entry, and approvals.
Assistant or client	Scaffold and local pass/fail tests; report results before deployment.
Millwork	Endpoint probe, approved routing, verdicts, receipts, usage, and cost.
Your endpoint	Candidate evaluation; keep credentials, traces, and datasets in your service.

The deployment owner approves deployment. The endpoint owner approves test traffic. The evaluator-account owner approves evaluator spending. An authorized person approves each paid Millwork run.

Choose public or protected access

Access	Use when
Public	Anyone who can reach the URL may submit a candidate and read its verdict.
Protected	Content or verdicts are private. Enter the key on Millwork’s private page, never in chat, a URL, a receipt, or a command argument.

Connection and testing do not approve a paid run.

Keep control when an assistant helps

Give an assistant the recipe and non-secret rules. Ask for local results before deployment; enter credentials yourself. Approve endpoint test traffic and evaluator cost before they occur, then inspect the preview before each paid run.

Output checks

Agent workflows

Review jobs

Operations recipes

Choose an output-check recipe

Choose a recipe

Try an included example

Who does what

Choose public or protected access

Keep control when an assistant helps

​Choose a recipe

​Try an included example

​Who does what

​Choose public or protected access

​Keep control when an assistant helps

Choose a recipe

Try an included example

Who does what

Choose public or protected access

Keep control when an assistant helps