What we do
We measure the gap between saying and doing.
Most benchmarks ask whether a model does what it's told. We ask whether it stays honest, fair, and safe while doing it. For each of ten virtues, we record what a model says about itself, then test what it does. The distance between the two is dissonance, and it is measurable.
Why it matters
A model can describe good values and still act against them.
A single score hides where a model fails. A model can answer a direct question about fairness well and still treat two people differently when only a name changes. Measuring each virtue separately shows which values hold under pressure and which do not, so teams can fix the right thing.
How we work
Evidence first, limits stated.
- Evidence. Every claim in a report connects to something we measured, with examples from the model's own replies.
- Independence. Answers are scored by judges from different model families than the model under test.
- Honesty about limits. Every score carries a range, and every report says what it does not cover. Our score bands are provisional while they are calibrated, and we label them that way.
The full procedure is on the methodology page, and a sample report shows what an audit delivers.
Where we are
Early, and building in the open.
The Alignment Audit is available, although it is still in development. Continuous monitoring, the Framework License, is coming soon. We would rather say what is still being established than overstate what is finished.