Reject misaligned observations in Model QA discrimination reports

[rudycelekli]

Agent Information

Agent file: specialized/specialized-model-qa.md

Motivation

AUC consumes the arrays positionally while pandas class masks align by index: reordered scores produce two inconsistent metrics in one report.

Testing

• Reproduced the original example failure; the corrected example passes the same focused fixture.
• Changed-agent lint and git diff --check pass.
• Full scripts/test-convert-outputs.sh --drift=advisory passes all 32 checks over 282 agents and 15 tool formats. Expected content drift is advisory; the generated-output manifest is left to maintainers.

Native extracted example with controlled local inputs; no production system or model-effectiveness claim. No paid API call, real invoice, production data or actual model-effectiveness evaluation was performed.

Candidate listed before publication in #917: #917 (comment). This single-profile correction adds no tooling or generated outputs. AI assistance was used to investigate, implement and test it; the commit is signed and includes DCO sign-off.