Three reports you can open, read and check yourself.

What it showsOutcome judgments in a buyer's words, over real tau2 airline transcripts: N of M per outcome, with each row's evidence class shown (fact, judged, or not yet showable).

What it doesn'tIt says nothing about EU AI Act or other obligations; that is example 2. A “pass” is a judged verdict against a pinned rubric and nothing more.

What it showsThe format we deliver: outcomes, daily AI grading, weekly blind human check, drill-downs, laid out over one month for an airline support agent.

What it doesn'tThe volumes and the weekly human checks are simulated to show the format. It is not a customer's report, no human agreement figure has been measured, and it is not sealed. For the sealed report, open example 1.

What it showsAll 37 rows of the EU AI Act Obligations pack, run over the same airline sessions as example 1. 34 rows read “not present” because a generic airline transcript does not carry the field the obligation needs. 3 judged rows read “not checked” because no judge was run.

What it doesn'tNo row reads established or failed. This is a scan of what the record can support, not a determination under the Act, and it is not legal advice.

What it showsHow two mesh-llm nodes each keep their own book of one paid exchange: the payer's payment events (merged upstream in #2035), the provider's matching half (proposed in issue #2017), and how the two join on the payment hash both wallets share.

What it doesn'tThat the inference was correct, which upstream does not check. It also doesn't show the swapped halves or the Evidence tab working today: those are not yet published, and the page marks each one.

What it showsQuadX AI's FCA obligations mapping (22 July 2026) for a buy-now-pay-later credit agent. It has 11 rows, and each is graded by who can check it: only the operator, a party to the transaction, or an independent third party. Rows 4 and 5 are backed by a real digest that two parties recomputed independently.

What it doesn'tThe other 9 rows are mapping only; they have not been run against any data. This is different data from examples 1 and 2, and it is not legal advice.

What it showsA buyer's agent negotiates with two competing seller agents on NANDA Town. A person clears the acceptance, and the agreed terms are sealed into one record per side, over the same digest. The records are designed so that anyone can recompute that digest and check both, once the service is public.

What it doesn'tNo payment moves, because the example stops at the agreement. The room seals both records; each side does not seal its own. The deal-room service itself is not published, and the page says which parts are merged and which are not yet public.

What it showsA buyer agent haggles with two competing supplier agents over three rounds and narrates its reasoning before each move. It picks a winner against a private mandate, and turns down the cheaper final offer because that offer breaks a hard constraint. Each side keeps its own signed record, and a stranger who trusts neither can check the whole transaction offline.

What it doesn'tNo payment moves. The agents' reasoning is scripted, with no model call. It is demo code that runs on your machine, not a hosted service or a supported adapter.

What it showsThe rules a judged row follows. The judge is pinned and sealed with each verdict, and the rubric is frozen before the period. Recomputed and judged rows are never blended, several judges can read the same row, and a blind human sample is reported as an agreement figure.

What it doesn'tIt gives no agreement figure for these examples, because none has been measured yet. The outcomes report's human check reads “no human-check data”.