fix(judge): improve 'poor' tier description to focus on errors, not missing data

[archasek]

Summary

The poor quality tier description in AxJudge reference-free mode currently reads:

poor: Response has significant issues or missing elements

This causes false poor classifications in data extraction tasks where empty/null output fields are correct — the source data is genuinely absent. The judge interprets correct empty fields as "missing elements" and downgrades to poor instead of good.

Fix: Change the poor tier description to:

poor: Response contains fabricated, incorrect, or contradictory information

This aligns the tier with actual quality problems (fabrication, errors) rather than penalizing structurally sparse but correct outputs.

Impact

• 1 line changed in src/ax/dsp/judge.ts
• No API changes, no new options, no breaking changes
• Default behavior for all other tiers unchanged
• Extraction eval tasks see ~5-8% score improvement by eliminating false-poor classifications

Test plan

• Tested on 25-case procurement data extraction eval
• Build passes

[Copilot]

Pull request overview

Updates the reference-free rubric in AxJudge so the poor tier penalizes fabrication/incorrectness/contradictions rather than “missing elements,” which is important for extraction-style outputs where empty/null fields can be correct when source data is absent.

Changes:

• Adjusts the poor tier description in the reference-free quality rubric to focus on incorrect/fabricated/contradictory content.
• Leaves all other tiers, scoring, and APIs unchanged.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.