TarskiTarski/ docsJoin the waitlist

Runtime containment

How Tarski keeps runtime LLMs in a bounded role, using relay namespaces, decision gates, invariants, and coordination through a shared derived model.

At runtime, Tarski treats LLMs as fallible tools behind relay surfaces. A model reads unstructured input and supplies candidate facts and proposed actions, and only your rules can accept them. A model never holds program control. This page describes the whole containment model.

The two relays

Model and sensor output lands in reserved relay namespaces:

  • candidate.*: fallible fact-domain evidence, such as an extraction, a parse, or a sensor reading. It cannot support a trusted fact until a decision.accepted.* rule that you wrote (and can inspect) endorses it.
  • proposal.*: fallible action-domain evidence, such as a drafted reply, a suggested offer, or a requested read. It cannot support a world-facing intent until decision.approved.* endorses it.

The loader enforces this as a relation-level graph cut over your whole program. If a rule wires model output directly into truth or action, the program does not load. A parallel bypass path fails the load even when a correct gated path exists beside it. See Facts, decisions & intents.

What a bad suggestion can and cannot do

A bad parse stays staged as a candidate. A bad proposal commits as evidence and derives decision.rejected.*, which is an auditable denial and is not treated as an error. An absurd order for eighteen thousand waters is recorded with provenance and cannot reach the point-of-sale. Tarski lets a model make a bad suggestion and keeps that suggestion from becoming a world-facing action.

Invariants add a fail-closed check on top of the relays. Declared impossible states, such as an effect receipt without canonical approval or an offer below the floor, are checked over the completed model at every fixpoint. A transition that violates one is refused with a receipt naming the invariant.

Reads are contained too

An agent that wants to read the worldview goes through the same process. It names a declared query, decision rules ratify or deny the request, an admitted intent executes the read at a pinned head, and results re-enter as evidence that is itself gated. Result rows cannot become trusted facts or drive intents without decisions, and this applies even to the bare fact that a query returned something. Budgets are ordinary decision rules with readable denial reasons, and invariants make over-approval an impossible state. For the full loop, see the agent query relay.

Curated context

What a model sees is declared, and nothing reaches its context undeclared. Every model-bound interface names a projection or an explicit full-worldview grant, and read evidence is excluded either way. A missing declaration is a load error. Context is assembled from digest-stamped sections whose digests join the provider request digest, so a capture envelope proves byte-for-byte what the model saw. You can find out what an agent knew when it made a proposal by reading recorded state.

llm.complete is one-shot. The request envelope has no tool surface, and capture validation rejects tool-invocation responses. A model that wants data asks through the query-request section, and the read happens as a governed effect on the next hop. There are no in-call tool loops.

Coordinating multiple agents

Multiple agents coordinate only through the shared derived model. Each agent contributes evidence, the evaluator derives one consistent worldview, and each agent reads through its own declared projection. Session isolation keeps simultaneous users separated by derived authorization facts. See Identity & tenancy.