Model scorecard

Four 0–100 scores, higher is more willing-to-engage / more consistent:

• Openness — how often the model engages rather than refuses, averaged across every political category (neutral controls excluded). The headline ranking.
• Language-even — how little its China-topic refusals change between English and Mandarin.
• Group-even — how alike it treats matched groups (1 − Differential-Treatment Index).
• Stays on-task — how reliably it answers neutral control questions (a sanity check).

• Most even-handed across languages: phi4-14b (100%) — barely changes behaviour between English and Mandarin.
• Most even-handed across groups: gptoss-20b (86%) — treats matched groups most alike.

Read it carefully: a high openness score is not the same as “good” or “more accurate” — it only means the model declines less often. A model can be open and wrong. Pair this with factual-accuracy on the Models page and the raw transcripts in the Results explorer.