Technology · Executable knowledge infrastructure

From source text to reproducible results.

Arxo turns laws, standards, textbooks and game rules into versioned, executable models. Your application supplies facts and a question. Through the SDK or MCP, it receives a result with the rules, source passages and edition behind it. One model can serve many applications.

Arxo executes the canon.

one question · one documentsdk · mcp
  1. Pinned sourceThe official text, pinned by a hash of its bytes. Every passage keeps its address.
  2. Versioned modelRules written against those passages, with editions and effective dates, published as a version.
  3. Facts and questionYour application states the facts and the date the rules apply on.
  4. Result and groundsThe result, the rules and passages behind it, the edition used, and a hash of the whole document.

Inspect the grounds. Reproduce the result.

Grounds

Trace the result to its source.

The answer is a document: the result, its status, the derivation from facts through rules to source passages, and the edition used. “The source is silent” and “this needs a person’s decision” are ordinary statuses with their own grounds.

status:   TRUE
result:   400 EUR
grounds:  Arts. 5(1)(c), 7(1)(b)
edition:  in force · 2005-02-17
Exactness

Calculate with exact values.

Money and quantities are exact numbers with a declared scale; rounding happens only where the source names a rule. Nothing outside the facts and the model can move the answer: no clock, no environment, no random source.

0.1 + 0.2   = 0.3
7 / 3       → refused: no rounding rule
Reproducibility

Preserve the inputs. Verify the output.

Ask on a date and the model applies the edition in force on that date. Keep the model version with its dependencies, the facts, the question and the date, and the run can be repeated; the hash of the canonical document confirms that the result is the same.

model:    eu.transport.air_passenger_rights@0.1.0
facts:    distance 2,300 km · notice 5 days
date:     2026-03-01
document: sha256:1c9e…4b7a

Cross-checked. For the rules on vehicle damage assessment in Kazakhstan, a model written independently in Catala agreed with Arxo on all 305 test cases.

Two readings, one document. Nine votes in the UN Security Council, one permanent member abstaining: the text and later practice give different answers, and both are laid out with their grounds. Open the published result on Lens ↗

How a source becomes a reusable model.

A model is written once against its source and then shared by every application that needs it. When the source changes, the change arrives as a new version of one dependency.

  1. Pin the sourceThe text is fixed by a hash of its bytes.
  2. Write the rulesEach rule points at the passage it comes from.
  3. Check with scenariosThe package carries its own test cases.
  4. Publish a versionA lock file fixes every dependency.

Update the rules. Keep the model.

Deploy changing rules without retraining your model. When a rule changes:

Fine-tuning
The rule lives in the weights: prepare new training data, train again, evaluate the new model.
Retrieval
The model reads the new text and applies it on its own; how it applied the rule is not recorded as a program.
Arxo
Publish a new version of one package. The model stays as it is; the result carries the rules, passages and edition behind it, and earlier results replay on the edition they used.

Verifiable training signals. For the same facts and edition, a canon computes the expected result with its grounds, so it can score a model’s answer as well as serve it.

Models formalize; two engines execute and prove. Authors work with the language guide, the editor extension (Marketplace · Open VSX) and the law CLI.

Choose your integration route.

MCP

Let your AI agent query executable models and inspect the grounds of each answer. Connect an AI agent →

claude mcp add --transport http \
  arxo https://mcp.arxo.io/mcp

Generated code

Arxo generates standalone source modules that return an answer and its status, in 16 programming languages and as Excel workbooks. Build and run them with your own toolchain.

Discuss code generation →

Compare changes. Find unintended outcomes. Map a dispute.

impact

What a change moves

Run a set of cases against the current and the proposed version of a source and see which results change, and through which rules.

loophole

An outcome nobody intended

Search the space of facts for combinations that reach a result the authors would not have wanted, with the derivation attached.

argue

The map of a dispute

Where a passage can be read two ways, both readings are laid out with the facts each needs, so the point of judgment is visible in advance.

Screening, procedures and plain-language explanations are described in the documentation →

Put your source to work.

Bring a rule, formula or requirement from a source you use. Start with one calculation or check, then choose how to connect it to your application.