The finding

1. Five yes/no questions: loops · chooses · acts · recovers · unsupervised. 4–5 yes = AGENT, 2–3 = WRAPPER WITH AMBITION, 0–1 = CRON JOB WITH EXTRA STEPS.
2. Three axes kept apart: capability (yes/no/uncertain), proof (how directly it can be checked), and leash (a measured number). A missing trace lowers the proof grade, never the capability.
3. Verdicts come from vendor documentation, read by AI research agents and adjudicated by hand against the published rules.
4. Every criterion a verdict depends on was reviewed individually before publication.
5. Frameworks are scored on what they do out of the box — a category distinction, not a product failure.
6. Anyone can overturn a verdict with evidence. Cases and rulings are public in the court.

We publish these because a rating that never moves is a rating nobody checked. Each row is a verdict we got wrong and the evidence that changed it.

The full dataset — every criterion, its result, reasoning, source, proof grade, confidence and check date — is open and CORS-enabled.

dataset.json dataset.csv rss social card

If you want to write about this, take the text. Each block is filled from the live dataset, so it cannot quote a number this site no longer stands behind. Nothing here is posted or emailed by any code we run — that is a human action, taken by a human.

Vendors: if your product is in here and you think a criterion is wrong, the court is the front door, and a vendor's own filing is labelled as such rather than buried.