Skip to content
← All research

AI & formal methods / Research prototype

What happens when Jev meets Lean?

A probabilistic model can propose a decision. A formal evaluator can make the acceptance rule explicit and reproducible.

The idea

Suppose an AI reads a patient note, a claim, a construction description or a tax record. It proposes a structured interpretation. Before software acts, a second layer checks exactly which facts support that interpretation, which policy applies and whether enough information is present. This is the architecture we are exploring across five open repositories.

The result can be deterministic without pretending that language understanding is certain. Freeze the candidate, evidence, thresholds and rule revision, and an ordinary total function returns the same answer every time. A changed model response may produce a changed decision. Saving the response is part of making the decision reproducible.

What Jev contributes

TypeSafe describes Jev as a System One model: state and typed questions go in; structured decisions and probabilities come out. Its public description names a parallel sampler and Reinforcement Learning for Calibrated Decisions. This is useful for decomposing an ambiguous workflow into small judgments.

The public material does not disclose enough to reproduce Jev’s complete architecture, training corpus or training procedure. Our release implements an independent verification layer and synthetic examples. It does not train or reproduce Jev. A value fitting its declared type also needs factual support before it can justify a real-world action.

Where Lean enters

Our bridge makes the acceptance boundary explicit. Candidate fields are checked against structured evidence and a fixed policy. Missing evidence, inadequate probability support or an unsupported case goes to review. Model probabilities are routing signals; they cannot supply a missing fact. TypeSafe’s separate confidence field is validated as metadata, without being treated as independent evidence.

The theorems establish properties of this encoded evaluator. They do not establish that an external model read a sentence correctly or that a policy author captured every applicable rule. That boundary is especially important when two apparently independent questions are evaluated against the same ambiguous source.

The connection to Lexifina

Lexifina’s case-law article describes a system in which extracted claims carry certificates tied to source sentences, and Lean checks encoded authority conditions. Its central design move is relevant here: the reader proposes; the checker enforces the formal acceptance criteria.

Our work applies that pattern to small, executable decision models. We have not reproduced Lexifina’s corpus or independently verified its reported results. The contribution here is inspectable code and a concrete way to ask what a decision depends on.

What comes next

The next useful experiment is a held-out, expert-reviewed corpus with frozen rule revisions and saved model responses. Measure error among accepted cases alongside the proportion sent to review. Report coverage, factual extraction errors and rule-encoding errors separately. Synthetic tests establish behavior at known boundaries; they cannot estimate production accuracy.

A Lean-backed system is valuable when its result can be replayed, its assumptions can be inspected and its uncertain cases remain visible. That is the deterministic result this release begins to make concrete.