TarskiTarski/ docsJoin the waitlist

Facts, decisions & intents

How Tarski turns fallible evidence into trusted state and authorized action through candidates, proposals, the decision namespace, and the graph-cut safety boundary.

Inside the evaluator, candidates, proposals, decisions, trusted facts, intents, and inspection relations are derived tuples in one fixed point. The taxonomy on this page is the vocabulary Tarski enforces at load time, and it is simple enough to use in a product review.

The two domains

Derived relations belong to two product-facing domains:

Domain Provisional form Committed form Question answered
Fact candidate.* trusted app-defined fact What is true?
Action proposal.* intent.* What should happen?

The two domains follow the same pattern:

candidate : fact :: proposal : intent

candidate.* is provisional evidence about state. proposal.* is provisional evidence about action. Trusted facts and intents are both committed outcomes, but they are not interchangeable. A committed fact does not trigger an effect by existing; only a committed intent may be admitted to an effect.

Decisions

decision.* is the deterministic adjudication between provisional and committed. If you replay the same observations and rules, Tarski derives the same decisions. Canonical decisions use this form:

decision.<outcome>.<domain>.<subject>(...)

The outcome segment is a closed set, and the loader checks it:

Outcome Domain Meaning May authorize
accepted Fact Candidate evidence is accepted as trusted truth Trusted facts derived from fallible evidence
refuted Fact Candidate evidence is denied as trusted truth Nothing (auditable denial evidence)
approved Action Proposal evidence is authorized for the reviewed action Reviewed world-facing intent.*
rejected Action Proposal evidence is denied for the reviewed action Nothing (the proposal remains evidence)
requires_review Either The subject must be routed for human or agent review Review-routing work only

decision.requires_review.* declarations must expose reviewer_kind: text and reviewer_id: text. Review is not a terminal authorization. Reviewer output re-enters the lineage as observations, and later rules derive accepted, refuted, approved, or rejected when the app has enough evidence.

The graph-cut safety boundary

Tarski enforces the decision model as a relation-level graph cut over positive dependencies from fallible relay relations (candidate.*, proposal.*, and runtime-populated rows such as queryresult.<query_name>):

  • Every positive dependency path from fallible relay evidence to a trusted fact must pass through decision.accepted.*.
  • Every positive dependency path from fallible relay evidence to a reviewed world-facing intent must pass through decision.approved.*.
  • decision.requires_review.* can cut only a path to an explicitly manifest-tagged review-routing intent. It cannot authorize the reviewed action.

The loader checks this as a property of the whole dependency graph, across rules. If a fallible relation reaches a trusted fact through an intermediate relation, the path is still fallible-origin until it crosses the right decision outcome. A parallel ungated path fails the load even when a correct gated path also exists.

Decision relations must carry the payload that downstream facts or intents need. A rule that derives trusted.fact(id, detail) from decision.accepted.foo(id) and then joins back to candidate.foo(id, detail) still bypasses the payload decision. Carry detail on the accepted decision relation instead.

Trivial acceptance is allowed, but it must be explicit:

decision.accepted.identity.verified_email(addr) :-
  candidate.verified_email(addr).

That rule states in the ontology, with provenance and fixture coverage, that this app auto-accepts this candidate shape. The promotion is explicit and visible.

Effects and invariants

Effects are the operational end of the action domain:

proposal.* → decision.approved.* → intent.* → admitted effect request → attempt → receipt observation

The fact domain carries epistemic complexity before commitment: unknown, stale, refuted, contradicted, or review-required evidence. The action domain carries operational complexity after commitment: admission identity, capability, attempt state, retry policy, receipts, and reconciliation. Tarski keeps epistemic complexity in the ontology and execution complexity in the runtime.

Invariants are fail-closed backstops over the fully materialized model. Handle ordinary bad proposals with decisions: a bad proposal should commit as proposal.* evidence and then derive decision.rejected.* or decision.requires_review.*. An invariant should reject the attempted checkpoint only when the completed model violates a declared impossible state.

Authoring checklist

  • Treat observations and typed atoms as evidence input; they are not trusted facts.
  • Put fallible sensor or extraction output under candidate.*.
  • Put fallible action suggestions under proposal.*.
  • Use decision.accepted.* / decision.refuted.* for fact-domain outcomes.
  • Use decision.approved.* / decision.rejected.* / decision.requires_review.* for action-domain outcomes.
  • Carry every fallible field needed downstream on the decision relation.
  • Mark review-routing intents explicitly in the manifest. Naming alone does not mark them.
  • Derive reviewed world-facing intent.* only from decision.approved.*.
  • Use invariants for impossible committed states, and use decisions for ordinary denial.
  • Prove happy, denied, review, contradiction, and bypass paths with fixtures.

Common mistakes

Mistake Why it fails
Deriving a trusted fact directly from candidate.* Missing decision.accepted.* graph cut, so the loader rejects it
Deriving an executable intent from proposal.* Missing decision.approved.* graph cut, so the loader rejects it
Letting decision.requires_review.* authorize the reviewed action Review only opens review work
Joining back to fallible evidence after a decision for payload fields The decision did not endorse the full payload
Treating rejected proposals as invariant failures Ordinary denial should commit as decision evidence
Calling an intent a fact Fact is reserved for committed fact-domain truth