- Each outcome and each obligation, as n of m sessions, marked met, not met or not evaluable.
- Recomputed rows kept apart from judged rows. A person checks a blind sample of the judged rows, and the agreement is stated.
- The sealed records each row rests on, verifiable offline, with the month's checkpoint on a public witness.
For enterprises
An Evidence Report for consequential work.
When an agent makes decisions that matter to your customers, your auditors or a counterparty, a dashboard is not enough. An Evidence Report is a monthly report built from sealed records of what your agents did. It is judged against Terms Packs you can read, and whoever you hand it to can check it without you.
- It gives no single percentage for the whole agent.
- It gives no verdict on the law. The obligation rows say what the record supports, and the report is not legal advice.
- It doesn't hide the unflattering rows. Declines, refusals and work nobody assessed are counted.
See what one looks like: three worked examples, including the monthly obligations report from the Amsterdam talk. The rows come from the packs.
How the rows read
Every row says met, not met, or not evaluable.
There's no percentage and no single number standing in for the month. Each outcome or obligation gets a count in each of three columns, and a reader can open any count down to the sealed records behind it. The calibration row says how often people, reviewing a blind sample, agreed with the judge.
This table shows the layout only. The letters stand for counts. They are not results.
| Outcome or obligation | Met | Not met | Not evaluable | How it was decided |
|---|---|---|---|---|
| Resolved the customer's request in one interaction | a | b | c | judged by a pinned model, method named in the report |
| Refund stayed within the published policy | d | e | f | recomputed from the record |
| Calibration: blind sample checked by people | k of n sampled cases where the people agreed with the judge | people, reviewing blind | ||
Open one
The sample report.
Stand-in. The independently checked monthly sample is not published yet. Until it is, the sample here is the sealed EU AI Act example. It is an obligations scan, not a monthly report: its rows read “not present” or “not checked”, it has no calibration sample because no judge was run, and its stamp reads “Countersigned: none”.
The report is one self-contained file. Its last section re-runs ten checks on the sealed records inside it, in your browser, offline. Change one byte of the file and the checks fail.
Open the sample report → · 0.2 MB
When a report has a second signature, its stamp shows it. What a countersignature adds →
Start with records you already have
See what a report shows on your own past month.
We run the same report over a past period of your agent's records, so you can see what it shows before you change anything.