Take it and check us.

• 30products
• 150criteria
• 150/150human-reviewed
• 2uncertain
• 28confidence below 7
• 32revisions on record
• 6court cases

1. This is not a model benchmark. It evaluates shipped products with tools, permissions and an operator — not foundation models. Two products on the same model can score differently, and usually do.
2. Documentation is not a trace. The proof grade on every criterion says which one decided it. 16 criteria rest on the vendor's word alone, and are labelled DOCS CLAIMED wherever they appear.
3. Leash comes from one sample. 42 Claude Code sessions · one operator · median 15 · max 92. It is a scale to read your own number against — not an industry benchmark, and not a claim about any other product or person. No contributed sessions yet.
4. Capability is a judgement, made in public. Every criterion carries its reasoning, its source, a confidence and a date, and every change to one is on the record. 32 verdicts have already moved.
5. Uncertain is not a soft no. 2 criteria could not be decided either way on the available evidence. They score zero and are shown as UNCERTAIN, never folded into NO.
6. Vote totals are not a survey. They are self-selected visitors, hardened against stuffing but not representative of anyone.

isitanagent.com. "Is it actually an agent?" 1.0, 2026-08-11.
25 of 30 shipped AI products pass a five-criterion agent test.
Dataset: https://isitanagent.com/api/dataset.json (CC BY 4.0).

Corrections are the point: file evidence in court and the dataset changes with your reason attached.