Model preference by study
- Original base model
- Fine-tuned model
Percentages use only votes choosing one model. Neither, skips and historical ties are reported separately below each bar. Earlier ratings remain included in their own dataset. Rows are never pooled.
A small pilot, not a general benchmark. Sessions identify browser tabs, not unique people; each session can rate multiple examples. More independent raters are needed, and preference does not establish factual accuracy. Records follow the existing 90-day session retention policy.
Separate automated assessment
AI review of the Methods drafts
Both versions of all 32 study briefs, checked for fidelity to the supplied methods and judged separately for manuscript style.