Support Ticket To Eval
Purpose
Turn real support pain into repeatable eval cases for product, support, or agent behavior.
Fit
- Use when support evidence should become a reusable regression or model-behavior eval.
- Do not use when the ticket only needs a one-off human reply and no future check is useful.
Inputs
- Support ticket, complaint, chat transcript, or failed assistant output.
- Product area, user role, and known expected behavior if available.
- Any sensitive details that must be redacted.
Workflow
- Redact private data while preserving the failure pattern.
- Identify the user goal, current behavior, expected behavior, and severity.
- Convert the case into a minimal input that reproduces the behavior.
- Write assertions that can be checked by a human or automated judge.
- Record variants that should be tested later.
Output
Produce Markdown with:
- Eval Scenario
- Source Evidence
- Test Input
- Expected Behavior
- Failure Mode
- Assertions
- Variants
Validation
- The eval preserves the original user pain without exposing secrets.
- Assertions are concrete enough to judge.
- Expected behavior is not invented when product intent is unknown.
- Severity is justified by evidence.
- The case can be reused without the full ticket context.