Markdown for LLMs
L4
The source Markdown for this article. Copy it into your assistant or download it as a text file.
# L4 **In short:** L4 and Arxo both turn a statute into named executable rules, and both were aimed at the same charities act. This page is for a reader who knows L4 — or its Jersey charities model — and needs one distinction kept sharp: limits of that published model are not limits of the language. ## What L4 is for L4 (Legalese) separates constitutive rules (declaring what counts as what) from regulative ones (parties bound by must, may, and must-not within time windows). Its strengths are the isomorphic shape — one provision, one named rule — typed deontics with fulfilment and breach tracked over event streams, an exception discipline written as rules guarded by negated exceptions, and tests living beside the rules with traces that visualize. Configured models deploy as rule services. ## Where it meets Arxo Both sides formalize the same act — the Charities (Jersey) Law 2014 — and answer the same charity-test, governance, and appeal questions with checkable derivations: Arxo's proof graph against L4's trace. The published Jersey model is a genuine third artifact here: it pins concrete readings of deadlines and amounts, some backed by secondary orders outside the statute, which the comparison treats as a separate edition-like axis rather than mixing it into the statute text. ## Key differences - **Evaluative terms belong to the model, not the language.** The published Jersey model computes words like "reasonable" or "analogous" from case data (non-empty strings, substring checks, list lengths). Arxo returns such questions to a named organ as requiring judgment. That is a choice of that model — the language itself offers other encodings, and the page does not attribute model shortcuts to L4. - **Exceptions are classical conjunctions.** An L4 exception reads as the rule body plus the negation of the exception, with no automatic priority inference or retraction. Competing grounds are structured by hand; Arxo asks for a declared priority and otherwise keeps the conflict visible. - **Tests explore; they do not assert by exit code.** L4's evaluation markers show values next to rules, but a run exits successfully even when a recorded assertion fails — recorded expectations are reading material, not a run report, until the pinned pair is actually executed. - **The cost of review is a hypothesis, not a measurement.** Whether L4 models are cheaper to accept against the act text is exactly the kind of claim the machine-author axis is designed to test. Until those runs happen, it stays formulated as a checkable hypothesis. ## A concrete scenario The prepared experiment sets twelve Jersey cases — the charity test, public benefit, an empty purpose list, analogy, the Jersey connection, appeal deadlines, spent convictions, risk levels — plus four edit cases adding secondary orders and a year-end parameter. L4 evaluation outputs and Arxo answers are judged against the statute text, under two explicit arbiters: the statute alone, or the statute plus the cited secondary orders. L4 also joins the shared explanation scenario, answering the same three questions — why yes, why no, and what is missing — under fixed judging criteria. ## Choosing and combining Choose L4 when the shape of the work is provision-by-provision rules with typed duties over event streams and a service endpoint at the end. Look to Arxo when evaluative terms must stay open as judgment calls, when competing grounds need declared priorities with visible conflicts, or when dated editions and source-anchored provenance carry the review. Combined, L4 can serve the operational rule layer while Arxo frames the same act for audit: what was decided, under which edition, on whose judgment. ## Evidence and open questions - Sources checked: September 2026 (engine and model repositories read on pins; unknown-value propagation, money details, working days, and norm revisions not found in the docs — recorded as open, not as absent features). - Studied profile: the L4 engine around late September 2026 plus the published Jersey charities model; no versioned engine release was pinned for execution. - Basis: confirmed by documentation plus a prepared protocol with frozen inputs; comparative run not performed. - Open: executability of the pinned engine-model pair, unknown-value behavior, and the review-cost hypothesis above. ## Sources and reproducible materials - Companion pages: [L4: the charities case bank](/comparisons/l4-charities-case-bank/) and [One charities text, five systems](/comparisons/jersey-charities/). - L4 engine and modelling guides: [github.com/legalese/l4-ide](https://github.com/legalese/l4-ide) - Published Jersey charities model: [github.com/ielegante/l4-jerseycharities](https://github.com/ielegante/l4-jerseycharities)