Rules your experts can read
Tarski rules are written in a small Datalog language for relations, rules, invariants, and relays. Datalog comes from the same logic-programming tradition as the rule languages of early expert systems. Tarski is not Turing-complete, which lets the platform guarantee that evaluation terminates and check how modules compose and what each rule costs before anything runs.
How LLM output reaches an action
LLM output lands in relay namespaces: candidate.* for extracted facts and proposal.* for suggested actions. No rule may derive an accepted fact or an executable intent from them directly. Every path has to pass through an explicit decision.* rule, and the loader checks the whole rule graph when the app loads.
The decision rule here is a bounds check. Orders within bounds are accepted, and everything else goes to a reviewer. Your coding agent writes this file, and tarski verify tells it whether the scenarios and invariants hold.
-- the voice parser is a fallible sensor: its parse arrives -- as candidate.order(item, qty) evidence -- acceptance is an explicit decision rule decision.accepted.orders.line(item, qty) :- candidate.order(item, qty), menu.bounds(item, max), qty <= max. order.line(item, qty) :- decision.accepted.orders.line(item, qty). -- out of bounds → route for review, never submit decision.requires_review.orders.line(item, qty, "reviewer", "order_desk") :- candidate.order(item, qty), menu.bounds(item, max), qty > max. -- invariant: checked at every fixpoint invariant pos_only_accepted_lines { never pos.ticket(item, qty), not order.line(item, qty) }
What the restrictions give you
Tarski stays inside well-understood fragments of logic. That lets the platform check properties of your system that it could not check in general-purpose code.
Bounded impact
The rule graph bounds what any change can affect. Recomputation, hydration, and review are limited to what a change can reach.
Modules that compose
Namespace boundaries are computed from the ontology and checked by the platform. Independently developed modules compose, or the merge is rejected with a verdict before deploy.
Known cost before running
The shape of a rule determines its worst-case cost. Tarski reports which rules fall outside the tractable classes, and why, before anything runs. Performance regressions are diagnosed from the rules.
Indexes and query plans are derived from the ontology and declared queries. Tuning can change constant factors and nothing else: every plan must reproduce the model and provenance that the reference evaluator derives.
Written by agents, read by people
You will rarely write Tarski by hand. Coding agents write it against your scenarios and invariants and run tarski verify until they pass.
When someone does read it, such as an auditor checking an acceptance rule or an expert signing off on an invariant, each statement is short and declarative.