October
Request a demo
Prompt Bias Testing
Mental health companion · Evaluated 18 September 2026

Luna bias results

Axis-level evidence from identical user questions tested under different demographic profiles.

01 · Published result0 cases flagged
0.0%Flagged rate
No cases flagged

No evaluated pair reached the review threshold.

Across 163 evaluated pairs and 3 demographic axes, the highest differential score was 2/10. Scores of 7 or more are flagged for human review.

163
Pairs evaluated
0
Cases flagged
3
Axes tested
2/10
Highest score
02 · Axis-level results1–10 differential scale
AxisCasesMeanMedianMaximum
Location
741.0112
Name & ethnicity
531.0212
Name & gender
361.0011
How to read this

The judge uses a 1–10 differential scale; 7+ is flagged for human review. Differences can reflect appropriate personalisation, including dietary or health needs. A low flagged rate does not establish fairness across all users.

134 pairs were excluded because their demographic changes did not alter the input sent by the application. Gateway privacy processing can further mask demographic cues. Coverage describes application inputs, not a guarantee that every cue reached the upstream model.

Incomplete cases

3 target pairs did not produce two usable responses and are excluded from the flagged rate.

Download evaluation evidence (JSON)