Unicode Logic Kit

unicode-logic-kit

A Python toolkit for logic with Unicode operators — *parse, transform, and reason about* formulas of classical first-order logic and well beyond it: modal, temporal, hybrid, many-valued, fuzzy, intuitionistic, relevant, second- and third-order, description, dependence/IF, substructural, and a range of further non-classical logics. On top of that sits infrastructure for NL→logic research: one small API a verification loop can drive, one uniform verdict type over a dozen provers, and an evaluation toolbox for scoring LLM-generated formulas against benchmark gold data.

from unicode_logic_kit import MSFLParser, is_valid

phi = MSFLParser().parse( "∀x (Human(x) → Mortal(x)) ∧ Human(socrates) → Mortal(socrates)") print(is_valid(phi)) # True

Renamed in 0.31.0. Up to 0.30.0 this package was unicode-fol-kit

(import unicode_fol_kit): first-order logic is one of the logics it handles,

not the whole of it. pip install -U unicode-fol-kit still works — it installs

this package and forwards the old import name with a DeprecationWarning — but

new code should pip install unicode-logic-kit and import unicode_logic_kit.

Nothing else changed its name, the UFK_ environment variables included.

Why this kit?

• One facade, seven verbs, built for LLM loops. api.parse_any (dialect auto-detection over Unicode / TPTP / LaTeX / Prover9 / SMT-LIB — never raises, records every parser's objection), api.check (well-formedness + signature conformance with did-you-mean suggestions), api.equivalent, api.prove, api.countermodel, api.repair (a diagnose→suggest→fix generator whose fixer callback is *your* LLM), api.translate (logic-to-logic via a comorphism registry). Every result has a JSON-compatible to_dict(), and the API carries an explicit stability policy.
• Logics as values, and a translation that carries its side conditions. FOL(MSFOL(f)) converts a sorted formula and keeps its side axioms (each sort is non-empty, each sorted constant lies in its sort) *with* the term, as a Sentence; api.prove adds them as premises itself, which is the difference between a proof and a spurious countermodel. Each of the nine registry edges declares what it preserves (comorphism.GUARANTEES), which side axioms its image needs and which options it reads; a composed path reports the weakest guarantee on it. There is no implicit coercion between logics on purpose — see Translating between logics for why a translation is not an upcast.
• An MCP server out of the box. pip install unicode-logic-kit[mcp], then python -m unicode_logic_kit.mcp (or, with nothing installed, uvx "unicode-logic-kit[mcp]" mcp) exposes the toolkit as thirty-seven Model Context Protocol tools (23 general-purpose, 6 for chemistry, 8 for description logic) — any MCP client (Claude Code/Desktop, agent frameworks, editors) can parse, prove, diagnose and translate without writing Python; the repair loop inverts naturally (the client LLM is the fixer). An error-analysis layer comes along: prediction-vs-gold breakdowns with symbol diffs (compare_formulas), corpus metrics (score_batch), satisfiability of premise sets with model witnesses (check_consistency), normal forms, cross-syntax rendering, truth tables, signature extraction, and DRS→FOL for discourse phenomena.
• Model checking against real structures, not just model finding. A molecule, a knowledge-graph neighbourhood, a scene graph is *given* — the question is whether a definition holds of it. FiniteStructure + evaluate_in_structure answer that directly: no prenex form, no CNF (both are documented timeout sources elsewhere), quantifiers ranging over indexed candidate sets rather than the whole domain, counting quantifiers counted instead of expanded, and properties that are decidable on a structure but not first-order definable over it (connectivity, ring membership) admitted as computed predicates. unicode_logic_kit.chem turns a SMILES string into such a structure over the ChemLog signature.
• Rule learning, both directions, with the silent failures made loud. unicode_logic_kit.ilp turns those same structures into an ILP task (Popper's bk.pl / exs.pl / bias.pl) and reads the learned Prolog clause back as a kit formula you can model-check. Two encoding mistakes produce a hypothesis that scores precision 1.00 and means nothing — example-local individual names a learner joins across, and the example argument on every predicate — and both are refused rather than documented, on the way out and on the way back. check_separation asks the question that has to come first: does your reference definition separate the two example sets at all?
• Exact probabilistic logic — no sampling, no floats. prob.entailment_bounds answers "what does P(bird)=0.9 entail about P(fly)?" with the exact tightest interval (Nilsson's probabilistic entailment as a rational LP over possible worlds, conditionals included); prob.query answers ProbLog-style queries over definite programs with independent probabilistic facts under Sato's distribution semantics — every result an exact Fraction, every unsupported fragment a loud refusal.
• Twenty-five prover backends, one honest `Verdict`. The kit's own calculi and semantic searches (resolution — with sound paramodulation/demodulation for equality, no hand-supplied congruence axioms needed —, analytic tableaux with recorded, independently checkable proof objects, modal tableau, finite model finder, bounded Kripke enumeration, QML embedding) are first-class citizens next to Z3, cvc5, Vampire, E, Zipperposition, Twee (equational proofs re-verified by an independent checker before the kit reports them), Prover9, Leo-III, Isabelle, nanoCoP-M (native first-order modal logic, opt-in with a mandatory independent cross-check) and a Dockerized HETS server (CASL export *and* round-trip import with native many-sortedness, multi-spec DOL library emission, SPASS/darwin behind one name, comorphism provenance in every verdict). Every route returns the same Verdict with a semantic status (proved / refuted / unknown / error), a separate *why-not-more* axis (timeout ≠ bound_hit ≠ incomplete ≠ unsupported), the SZS status, wall time, and JSON-able witnesses. Chains, portfolios (portfolio_prove, parallel with agreement thresholds and a soundness alarm on prover disagreement), and cached batch runs (batch_decide) are built on top.
• An evaluation toolbox for NL→FOL work. A graded equivalence ladder (exact → canonical → vocabulary-aligned → solver, with a tri-state solver level and partial credit), AST-level symbol alignment, gold-formula self-audit, and adapters for FOLIO, MALLS, GROVES, WillowNLtoFOL, ProntoQA, ProofWriter, LogicNLI, ProverQA and FraCaS — each with its upstream schema verified at the source and its limitations documented instead of smoothed over. FraCaS is the pure-NLI one: no gold formulas at all, so the translation step is an injected callable and the kit only decides.
• Countermodels that explain themselves. api.countermodel returns a machine-readable witness *plus* a plain-English rendering ("The countermodel has 2 possible worlds. … At world 0 the formula fails."), and refutation actually covers the temporal fragment: Ⓕ P → P comes back refuted with a two-world Kripke witness, not "unknown".
• Honesty as a contract. No route silently degrades: fuzzy input is refused by classical provers rather than collapsed, an unavailable backend raises instead of vanishing from the chain, incomplete methods report *why* they stopped, and every proof method has an independent checker.

from unicode_logic_kit import api

result = api.parse_any("∀x (Raven(x) → Black(x))") # LLM output, any dialect report = api.check(result.formula, signature={"predicates": {"Raven": 1, "Black": 1}}) verdict = api.prove(api.parse_any("Black(tweety)").formula, premises=[result.formula, api.parse_any("Raven(tweety)").formula]) print(verdict.status, verdict.backend, verdict.szs_status) # proved z3 Theorem

One parser class, MSFLParser, reads classical logic at the first, second and third order, each with or without sorts and with or without the modal family (modal/temporal/epistemic/deontic/hybrid): twelve combinations of three constructor flags. A fourth flag, fuzzy, puts the Łukasiewicz connectives in place of the two-valued ones in each of the twelve. Next to them stand team-semantic dependence/IF logic, intuitionistic linear logic, and the Lambek calculus. The surface syntax is natural Unicode (∀ ∃ ∧ ∨ ¬ → ↔ ⊕ ⊗ □ ◇ @ ⊸ 𝟙 …) with no ASCII fallbacks.

On top of the AST sits a full reasoning stack — four proof methods (a built-in resolution prover, Fitch natural deduction with checker *and* searcher, the Gentzen sequent calculi LK/LJ — the latter backed by Dyckhoff's G4ip, a genuine terminating decision procedure for propositional intuitionistic logic — and analytic tableaux), a finite model finder, SMT (Z3) and external-prover (Prover9 / Vampire) backends, truth tables, and dedicated semantics for every logic. Every proof method has an independent checker — including check_resolution_proof, which certifies externally produced resolution derivations (paramodulation steps included) with its own from-scratch unification, and check_tableau_proof, which re-derives every α/β/γ/δ step of a recorded tableau with its own substitution and δ-freshness check. Formulas import/export to TPTP (incl. header Status/Rating metadata via parse_tptp_problem), Prover9, SMT-LIB, LaTeX, and JSON.

📖 Documentation

Full guide and API reference: [https://unicode-logic-kit.readthedocs.io/](https://unicode-logic-kit.readthedocs.io/)

The documentation walks through every logic with runnable examples — start with the Quickstart and Choosing a tool.

Beyond deciding formulas: Model checking (evaluate a formula in a structure you already have — including molecules), Verification (is a definition *set* coherent, or just too easily satisfied?), Probabilistic logic (exact bounds and queries, no sampling), and MCP (the toolkit, grammar included, as tools for a language model).

Installation

pip install unicode-logic-kit

Requires Python 3.10+. Z3 ships with the package. pip install "unicode-logic-kit[cvc5]" adds cvc5 as a second in-process SMT backend (it then joins the default proving chain right after Z3); [mcp] adds the MCP server (python -m unicode_logic_kit.mcp); [hf] add