Every AI company publishes numbers. Most of them are selected, rounded, and framed until they say whatever marketing needed. We think that is backwards: benchmarks should be evidence, not advertising. Here is our policy, in writing.
What we publish
Full breakdowns, not just headlines. Our KB-100 result is not just 65.2%. It is five categories with baseline scores, gains, task counts, and a grading rubric, all on one page.
The unflattering parts, too. Constrained design went from 4% to 21%. That is still a low absolute number, and we published it anyway. Improvement on hard tasks tells you more when the starting point is visible.
Methodology before marketing. Model, effort level, task count, judge, and rubric are stated up front, before the claim.
What we will not do
Cherry-pick the one category that looks best and crop the rest. Compare our best configuration against a competitor’s weakest. Hide custom task sets behind vague names without describing them.
Why this matters for you
You are being asked to trust an intelligence layer with real work. Trust should come from numbers you can interrogate. Every benchmark claim links to the data behind it.
If you find a flaw in our evaluation, tell us. We would rather fix the benchmark than defend the score.