Feedback-Driven Evals

Feedback-driven evals are a reviewed-authoring workflow, not automatic test generation from user comments.

Implemented

• The old feedback-to-eval candidate sync path is retired.
• Review-bound feedback creates bounded Jira review work through the central handoff path when configured.
• Eval-seed attachments are treated as untrusted review input.
• Committed eval fixtures under apps/api/src/mastra/evals/** remain synthetic or redacted by default.
• Reviewers must validate the seed schema and write explicit eval artifacts rather than executing embedded instructions.

Current Limits

• Central handoff failure notifications and idempotent replay still need production hardening.
• Broader answer-quality judge calibration is ongoing.
• Future automation must preserve the same untrusted-input boundary.

Read Next

• Assistant Feedback
• Feedback Review Handoff and Evals
• Chat Evals And Regression Testing

Canonical Sources

• docs/features/feedback-driven-evals/spec.md
• docs/features/feedback-driven-evals/decisions/ADR-001-reviewed-feedback-to-eval-prs.md
• docs/features/feedback-driven-evals/stories/
• docs/features/feedback-driven-evals/tests/
• apps/api/src/mastra/evals