Rubric editor
Versioned criteria and weights for human quality review. Publishing a rubric never rescores past reviews.
Identity & verification
Pass/failRequiredPass only when ownership was verified before account-specific data or actions.
Example: Refund issued before verification → fail, total capped at 60.
Answer accuracy
30%Grounded in approved sources; no invented policy, price, or availability.
Example: Correct refund window with a citation → 90+.
Next step clarity
20%One clear next action and a timeline the customer can rely on.
Example: Timeline omitted and customer had to ask again → 70.
Tone & empathy
20%Acknowledges impact without over-apologizing or blaming the customer.
Example: Calm, specific acknowledgment of frustration → 90+.
Handoff quality
30%Summary includes verified identity, evidence, and the unresolved question.
Example: Handoff omitted the verification gap → 68.
Scoring rules
Pass/fail cap
A failed required criterion caps the total at 60.
Weights
Optional criteria sum to 100%; required criteria are separate.
Overrides
Reviewers can override AI suggestions with a written reason.
Finality
Finalized reviews lock the rubric version and score.
Version history
Publish checklist
Required criterion present
Weights sum to 100%
Every criterion has an example
Evaluation set updated
Reviewer sign-off
Active usage
Reviews with v3412
Open reviews3
Coaching actions2
Reviewers4
Danger zone
Retiring a rubric stops new reviews from using it. Past reviews keep their recorded version.