Methodology

How the dataset is produced

A fixed, versioned prompt matrix, run against pinned model snapshots rather than “latest”. Each answer is evaluated against published criteria. Raw results ship as CSV and JSON with a changelog; a changed prompt gets a new id and a new matrix version.

What this cannot prove

There is no control group. Where an article targets a set of prompts and that set later moves, the movement is reported alongside an untargeted remainder from the same matrix version — and that remainder is not a control group and is never described as one. Reporting a number and refusing to claim causation for it is a deliberate editorial position.