Pilot · Executable knowledge infrastructure

One source. One class of questions.

A pilot takes one authoritative text and one kind of question you already handle, and shows on your own material what an executable model of it returns: the result with its grounds — the rules applied, the passages of the source and the facts used.

There are three kinds. Each is fixed in scope, result, timeline and price before it starts.

01 Second check

A second check of a calculation you already run.

You compute payouts, contributions, fees or deadlines under a rule. We write an executable model of the same rule from its source and run your completed cases through it. Every difference comes back traced to its source.

You bring

The source your calculation follows: a law, a regulation, a standard or an internal method.
De-identified completed cases with the results your system gave.
One specialist who knows those cases, for a few hours a week.

You get

A versioned model of the rules, pinned to the source.
Every case run through the model, with its grounds.
A report of the differences, each traced to the rule, the edition and the fact behind it, with a status: defect in the text, different reading, error in one of the models, or no source.
Zero differences is also a result: a second check on record.

Report of differences · one entryIllustrative sample.

difference   07 of 40
case         C-0291 · cancelled · notice 3 days
your system  600 EUR
model        400 EUR · eu.transport.air_passenger_rights@0.1.0
rule         Art. 7(1)(b) · band 1,500–3,500 km
passage      Art. 7(4): distance by the great circle route
fact         3,412 km great circle · 3,560 km flown
edition      Regulation (EC) No 261/2004, in force since 17.02.2005
status       error in one of the models
             your system measured the flown route

Scope: One source, one class of questions.
Timeline: Usually 4–6 weeks; agreed before the start.
Who reviews: Your expert and a reviewer we name.
Price: We fix the scope, the result and the price before the start.
What we do not promise: We do not promise to find errors. The report is not a legal opinion and does not replace your system.

Propose this pilot

02 Draft or act

A consistency check of an act or a draft.

Before a draft is submitted, or while an act is in force, check that its duties, deadlines and references hold together on concrete situations. The work is private; nothing is published.

You bring

The act or the draft, and the acts it refers to.
The questions it has to answer: who must do what, by when, and with what consequence.
Typical situations from practice, if you have them, and a contact in the drafting or legal team.

You get

A private executable model of the act’s duties, deadlines and references.
The situations run through it, each with its grounds.
A card for every finding: the passages involved, the situation that shows it, the result and a status: defect in the text, different reading, error in the model, or no source.
No findings is also a result: a check on record.

Finding cardIllustrative sample.

finding      3 of 9
passages     Art. 9(1), Art. 9(4), Art. 14(2)
situation    application refused, then granted on review
result       two decisions; the 30-day deadline of Art. 14(2)
             runs from "the decision" and has two start dates
status       defect in the text
for          the drafting team: name the decision that starts
             the deadline

Scope: One source, one class of questions.
Timeline: Usually 4–6 weeks; agreed before the start.
Who reviews: Your expert and a reviewer we name.
Price: We fix the scope, the result and the price before the start.
What we do not promise: We do not promise to find defects. The cards are not a legal opinion and do not replace your drafting or review procedure.

Propose this pilot

03 Agent

Connect your agent to a model that answers with its grounds.

Your assistant or application answers questions that depend on a rule. Instead of paraphrasing the text, it asks a pinned model and returns the result with its grounds, or says which fact is still missing and when a person has to decide.

You bring

The agent or application, and the questions its users ask that depend on a rule.
The source those answers must follow. If no model of it exists yet, writing one is part of the pilot.
A sample of real questions, de-identified.

You get

Your agent connected over MCP or the SDK to a pinned model version.
Answers that come back with their grounds, the facts still missing, or a note that a person decides.
A record of each answer to keep with the transaction and run again.
If your stack needs it, a generated module in TypeScript, Python or Go that returns the answer and its status.

Answer record · recorded sessionFrom the recorded session on the MCP page, shortened.

agent → law_ask  compensation_due · 19 facts · date 2026-09-28
status           TRUE_ONLY · computed
result           400 EUR
grounds          Art. 5(1)(c), Art. 5(4), Art. 7(1)(b)
missing facts    none
for a person     none in this case
checksum         sha256:840fb48a…dcab03f4

Scope: One source, one class of questions.
Timeline: Usually 4–6 weeks; agreed before the start.
Who reviews: Your expert and a reviewer we name.
Price: We fix the scope, the result and the price before the start.
What we do not promise: We do not promise that every question your users ask is covered: the model answers its class of questions and says when a question is outside it. It is not legal advice and does not replace your product.

Propose this pilot Prefer to start on your own? Connect an agent over MCP

Propose a pilot.

Write a few lines: the source, the questions and the material you have. We reply with a proposed scope, and agree the reviewer, the timeline and the price before any work starts.

Write to hi@arxo.io See a result with its grounds