The Law Is Becoming Typed Input
Six papers crossed this survey's desk within a single month, and together they describe the same event from six directions: law is being rebuilt as typed, versioned, adjudicated input to inspectable decision systems — not as prose that a language model interprets on the fly, and not as model output to be trusted. The statute is becoming a data structure. The question is whether the structures now being proposed are honest ones: where the law is genuinely mechanical, do they mechanize it — and where it is genuinely open-textured, do they say so, in the type system, instead of guessing?
The six exhibits
Where Do Legal Factors Live in LLMs?
CATO legal factors — the structured issues lawyers reason over — turn out to be linearly decodable inside a language model, localized to a small set of sparse-autoencoder coordinates (78 layer-features). The machine already keeps its legal reasoning in identifiable places; this work finds the addresses.
Machina Pauliana
A statute rendered in Catala — a programming language built for law — with a neuro-symbolic wrapper: the mechanical parts compile, and the open-textured terms (the phrases no honest reading can mechanize) are exposed as typed inputs that a human must supply. The ambiguity is in the schema, not swept under the output.
Grammar of French Civil Cassation
The reasoning of France's highest civil court, described as a formal grammar: the moves a cassation judgment is allowed to make, in the order it may make them. Appellate review as syntax — a decision procedure you can parse, and therefore audit.
ARCCS
Regulatory compliance approached as a decompilation problem: take a regulation as shipped — dense, cross-referencing, amended — and recover the structured requirements it actually imposes, in a form a system can be checked against.
UCVE: Graph-Structured Retrieval for Compliance Validation
Compliance as a coverage graph: requirements are nodes, the evidence that satisfies them is retrieved along learned edges, and a spectral reranking decides what counts as covered. Validation becomes a question with a graph for an answer — inspectable, node by node.
GRACE
Statute grounding: the seam where statutory text meets a model, treated as the place where errors are born and therefore where engineering belongs. Grounding is not a detail of the pipeline; it is the pipeline's load-bearing joint.
What it means
Read together, the six describe an architecture with a consistent shape. The law enters as data — typed, versioned, cited to its source. A deterministic evaluator applies it; the model proposes, but the checker disposes. Where the text is open-textured, the system abstains upward to a human, through a typed slot — it does not improvise a permission. And every decision carries its trace: which clause, which version, which evidence, which human gate.
That shape matters most where the data is most sensitive. Healthcare runs on permissions nobody wrote down cleanly: data use agreements, consent scopes, retention clauses, role definitions scattered across contracts and policy PDFs. The promise of this literature is not that machines will "do compliance." It is narrower and better: that the distance between the agreement as signed and the permission as enforced can be made visible, measurable, and short.
Coda: the next wall is consent
The same week this post was drafted, the survey's net pulled four consent papers from four unrelated waters. Two are worth flagging now. A study of end-of-life digital privacy (arXiv 2610.25773) documents the pattern everyone who has managed a family's accounts already knows: as care expands around a person, their control over their own data quietly contracts — consent thins precisely when the stakes thicken. And EREBOTS 4.0 (arXiv 2610.36221) describes agent-to-agent systems in which consent flags travel as first-class context between machines — the permission is part of the message, not a policy PDF neither agent has read.
Consent as infrastructure — typed, carried, revocable, logged — is where this survey goes next.
The fields under survey
Forthcoming
- Consent as infrastructure — flags that travel with the data
- The de-identification seam — what "anonymous" survives contact with linkage
- Federated learning's quiet governance problem — who audits an update nobody saw?
Titles under survey, not promises. A post appears when the papers do.