Skip to content

Resources/Prepare

Quality method

Define useful quality checks

Measure the behaviours the task requires instead of hiding the decision behind one generic quality score.

Audience
Domain experts, product owners, and AI practitioners agreeing how responses will be judged.
Question
What must a good response do, and what failure would matter to the decision?
01

Turn requirements into separate checks

A response can be correct but unusable, fluent but unsafe, or well structured but incomplete. Keep distinct requirements separate so one strength cannot erase a blocking failure.

  1. 01

    Name required facts or actions.

  2. 02

    Name prohibited claims, tools, or remedies.

  3. 03

    Check the response structure when downstream code depends on it.

  4. 04

    Add tone or style only when it changes customer or reviewer usefulness.

02

Choose the right reviewer for each check

Use deterministic code for exact structures and values, a qualified person for ambiguous domain judgement, and model-based scoring only where its rubric and disagreements can be reviewed.

  1. 01

    Prefer an exact check when an exact rule exists.

  2. 02

    Record the human role needed for policy or domain interpretation.

  3. 03

    Test an automated judge against reviewed examples before relying on it.

03

Set blocking criteria before results arrive

Agree which failure prevents a change even if the candidate wins on average. This avoids moving the threshold after seeing a preferred result.

Worked example

Example: billing-dispute response

Required: explain the disputed charge and the approved next step. Prohibited: promise a credit without an approval state. The prohibited action is a blocking check; an otherwise polished response still fails.