Skip to main content
Human guide: work with your agent → Test an authored scoring rule against known accepted, rejected, and incomplete accounts. Review the errors before applying that rule to new accounts.

​
Before you start

Use an authorized workspace and a small file with stable source IDs. Have source access and your team’s acceptance policy available. This guide describes customer-authored logic, not an automatically enabled product policy.

​
1. Understand the decision

A score is an authored decision rule. Keep its version and the evidence used to compute it. Test this fictional sample before calling providers:

​
2. Inspect supported operations

Inspect a returned ID with deepline describe <returned-id> --json. Compare required identifiers, outputs, access mode, and price. A similar title does not establish that a contract fits.

​
3. Implement the bounded workflow

Ask your agent to author account-scoring-workflows.play.ts using Write Plays. Have it declare the CSV input and preserve source IDs. Reuse existing contracts when they fit the transformation. Implement these rules explicitly:
  1. Define inputs, weights or rules, exclusions, and missing-data behavior.
  2. Use accepted account evidence and preserve observation times.
  3. Test labeled fit, non-fit, and uncertain cases before adding providers.
  4. Return the score, contributing reasons, policy version, and review state without changing routing.
Keep external changes disabled for the first pilot. That control belongs in the authored logic, not a universal Play dry-run flag.

​
4. Check and pilot

Prepare pilot.csv using CSV-safe selection. Then check the source:
Resolve authoring errors first. After reviewing source and price, run only the pilot:
Use the ID returned by the pilot. Select a returned dataset path when several exist. Provider-backed execution can incur usage.

​
5. Verify the outcome

Compare calculations with the labeled cases. A model-written explanation must agree with the rule and recorded inputs.
Every source ID has a result or explicit exception. The fictional cases follow the stated policy, and actual results retain their evidence.
Ask your agent why a record was accepted, held, or rejected. Correct the rule before expanding input or enabling external writes.

​
Continue

Prioritization connects scores to actions. Write Plays covers custom executable logic.