Predicate evaluation contract

harness.llm.evaluate_predicate(id, question, input, policy) declares a probabilistic evaluation. Its frontend contract is available for checking and tooling. Execution currently refuses with an explicit unavailable-evaluator error. It makes no model request. Budget admission, receipts, replay, and model execution are separate implementation steps.

Import the public types from std/predicate:

import "std/predicate"

fn assess(
  llm: HarnessLlm,
  input: {claim: string, observation: string},
  policy: PredicatePolicy,
) -> PredicateOutcome {
  return llm.evaluate_predicate(
    "finding.support.v1",
    "Does the observation support the claim?",
    input,
    policy,
  )
}

Arguments#

The policy requires backend: "structured_llm", string fields provider, model, and effort, and floating-point fields temperature, threshold, evaluation_cost_limit, and run_cost_limit. Checking validates this shape; it does not resolve a provider or establish a resource reservation.

An unvalidated any, unknown, bare dict or list, open record, recursive type, function, or capability handle cannot be an input. Validate external data before calling the evaluator. Taking the evaluator method as a function value or calling it through optional method access is rejected; a typed helper retains the literal source site instead. The method receiver must retain its static type; erasing it to any or an untyped map does not bypass site checking. Dynamically computed property calls do not declare admitted sites. The current runtime refuses all evaluation, and the executor must require an admitted artifact site before enabling execution.

Outcome#

PredicateOutcome is a closed tagged union. Its kinds are verdict, low_confidence, refused, budget_cut, unavailable, replay_mismatch, and cancelled. Every variant has a receipt reference. Only verdict exposes an accepted value with verdict: bool, confidence: float, and evidence: string.

import "std/predicate"

fn disposition(result: PredicateOutcome) -> string {
  match result.kind {
    "verdict" -> {
      if result.value.verdict {
        return "candidate"
      }
      return "not supported"
    }
    _ -> { return "unassessed" }
  }
}

An outcome cannot serve directly as a boolean condition. Match its kind before reading its verdict. The checker rejects unused plain bindings, discard bindings, and discarded outcome expressions; returning an outcome transfers its handling to the caller. These checks establish explicit use, not whether the caller's policy is correct. Model judgments never establish type proofs or grant authority. Variant-specific property reads, indexed reads, and destructuring also require narrowing. Common kind and receipt fields remain available on every outcome.

Site manifest#

harn check --json includes files[].predicate_manifest. A successful analysis returns schema harn.predicate_sites.v1 with a sites array. Each site contains its ID, question SHA-256, canonical input-type SHA-256, outcome schema, declared effects, and source path/line/column. The census includes transitive imports, so checking an entry file also reports sites declared in its helpers. The check-result cache preserves the manifest and tracks imported source changes.

An empty sites array means the checked import closure declares no evaluation sites. null means the file failed checking and has no complete census. A site is a source declaration, not evidence that a model ran. Runtime inputs, credentials, and model answers do not appear in this manifest.

Diagnostics#

The design explanation defines the remaining runtime, replay, budget, and provider contracts.