What this is

Arxo is formalized law: a normative act rewritten in a language that executes. A question to it is not a search for similar text and not a paraphrase of the statute. It is a computation. The engine applies the act’s rules to the facts of your case and returns four things at once: a status, the rules that fired, the address of each provision in the official text, and hashes that reproduce the answer later.

That is the property the whole system is built for: an answer can be rechecked, not only read. Every claim has an address, every conclusion has a chain, every chain has a legal edition and the date it was projected on.

Arxo executes the canon. Models formalize; two engines execute and prove.

Want to see it before reading on: a computed answer with its proof is on lens.arxo.io, the Kazakhstan slice with worked cases is zan.arxo.io, and your own first question from a Node application takes ten minutes — Quickstart.

Five words

Canon + facts + legalTime + question → Arxo Law → answer, provisions, grounds, and the data to replay it.

Word	What it means here
Canon	a normative text pinned as bytes and formalized as executable rules; named with an exact version, e.g. de.bgb.fristen@0.1.0
Facts	what your case asserts, as predicate(args) in the canon’s own vocabulary; a fact you do not list is unknown, not false
Question (predicate)	a closed claim to establish, a value to compute, a period to end, or the positions that arose — asked in the same vocabulary
legalTime	the date the law is projected on: the same question on two dates may lawfully give two answers
Answer (evaluation document)	statuses, the rules that fired, the provisions behind them, hashes of the canon, the case, and the result — the bytes a reviewer replays

Path of an answer

Step	What happens	Where you see it
Act	official text is pinned as bytes and hashed	law_sources
Package	norms are written as language rules; articles remain addresses	law_rules
Question	a predicate + arguments + facts of your case	law_ask
Computation	the engine applies rules to facts at legal time	four-valued status
Answer	status, rules, provenance, hashes; on request, a proof graph	law_explain

There is no intermediate “understanding of the text” on this path. A norm is either formalized and then applied, or it is not formalized and then silent. There is no third state in which the system “roughly guessed”.

What you can ask

The kind of question is the kind field of law_ask:

kind	Question	Answer
truth	is this fact established	one of four statuses
collect	list the records	a list of values
calc	compute a quantity	the value and how it was formed
deadline	when a period expires	a date on the official calendar
positions	which duties arose from the facts	positions and their statuses

Analytics sit on top of that: how robust the conclusion is (law_argue), what changed between two legal dates (law_editions), what stages a process has and how a particular case will run (law_process, law_process_run). The live list is on Connect an AI assistant (MCP).

What this is not

This is not text search. law_search is a discovery layer: it answers “what to ask next” by returning addresses of predicates and rules. It establishes neither a fact nor a norm. Textual similarity does not mean applicability.

This is not a language model. The answer is computed by the engine and deterministic: the same inputs yield the same bytes. A model may wrap the answer in words — it cannot produce the answer, and you must not attribute to the system an answer it did not give.

This is not legal advice. Terms are on Legal and licensing.

The corpus is not all of law. Part of the acts is formalized, and inside an act — part of the articles. Silence means “not formalized”, not “the norm does not exist”. How deeply a given act is taken apart is shown by the completeness measure (law_measure): an article may be executable, merely anchored, or present as text alone.

Not every package is the law of a state. The corpus includes doctrinal and teaching packages; the catalogue marks them, and every answer names a jurisdiction. A package that formalizes doctrine or a problem set answers under its own rules and is not the law of any country.

An answer is always as of a date. Law is projected onto legalTime, and the same question on two dates may lawfully yield two different answers. The date the answer was computed on lives in its provenance — not a default setting, a part of the answer.

Two answers most often taken for a fault

“Not established” is an answer. The world is open: absence of confirmation is not refutation. Status NEITHER says that formalized law has neither one nor the other, and with it comes whyNot — candidate rules with the status of each premise. What was missing is visible there; you do not have to guess.

An error is also an answer. A tool refusal means the question does not make sense in this act (no such predicate, wrong arity, wrong type). Answering it from general knowledge of another country’s law is the worst thing you can do.

How to read the rest of the answer — the next page.

Where to go next

• How to read an answer — anatomy: statuses, whyNot, provenance, hashes, proof.
• Connect an AI assistant (MCP) — how to ask from your agent or editor.
• Connect: the editor — highlighting and diagnostics for .law in VS Code and Cursor.
• Why this can be trusted — what reproducibility rests on and what it does not prove.
• zan.arxo.io — the Kazakhstan slice of the corpus: what is already formalized and how it answers on worked cases.