Agent Evidence Sufficiency Engine

Evaluate whether supplied structured evidence satisfies an explicit contract before an agent proceeds. One claim per run, JSON input/output, no LLM calls or external evidence fetching.

Returns SUPPORTED, CONTRADICTED, or INSUFFICIENT, with evidence IDs, exclusion reasons, source-group counts, freshness and provenance assessments.

Supported v0.1 scope

This is the deterministic exact-value subset of C.I. #3 freeze specification v2.0. It evaluates observations already normalized by the caller. It does not interpret prose, authenticate sources, verify real-world truth, independently establish source independence, or calibrate model confidence. No semantic evaluator or arbitrary predicate DSL is included in this release.

Input contract

• mode: precondition, trigger, claim, or outcome. A workflow label; decision rules are identical.
• claim_or_condition: { "id": "shipment_delivered", "expected_value": true }.
• evaluation_time: explicit ISO8601 timestamp with timezone.
• evidence_requirements: required minimum_independent_sources (1–1000), required_source_types (unique strings), max_age_seconds (nonnegative), confidence_threshold (0–1), require_provenance (boolean).
• evidence_bundle: up to 1000 objects containing unique id, claim_id, source_id, source_type, observed_at, and value. Supply confidence (0–1); missing confidence is excluded, not invented. Optional independence_group defaults to source_id. Optional provenance must contain a nonempty reference.
• resolution_policy: insufficient_on_conflict (default) or contradiction_veto.

Evidence is eligible only if its claim_id matches, its time is not in the future or too old, confidence meets the threshold, and required provenance is present. Eligible value is compared with expected_value using strict JSON equality, without coercion. A different value counts as contradiction: normalize observations before calling; do not use unrelated descriptive text as values.

Counts use unique independence groups, not evidence records. Different source IDs with a shared upstream origin must declare the same independence_group. The engine trusts these declarations; a copied record with a falsely independent group cannot be detected automatically. Repeated source IDs must declare consistent grouping. Source-type coverage and independent-group count are separate requirements.

A side must meet both independent-source count and required-source-type coverage. Pure support meeting the contract yields SUPPORTED; pure contradiction meeting it yields CONTRADICTED. Mixed eligible evidence defaults to INSUFFICIENT. Under contradiction_veto, any eligible contradiction blocks support, but CONTRADICTED still requires the full contradiction contract.

confidence is the minimum caller-supplied confidence on the decisive side, or null for abstention. It is not a calibrated probability that the claim is true. Provenance checks verify reference presence only.

Unknown fields are rejected to avoid silently ignoring unsupported requirements. Input errors fail the run rather than producing a verdict.

Example

See examples/input.json: delivery asserted by a carrier and acknowledged by an independent receiver, both fresh and with provenance references. Result: SUPPORTED. Remove the receiver: INSUFFICIENT. Change the receiver value to false: INSUFFICIENT under the default conflict policy. Change both to false: CONTRADICTED.

Output and use

The default dataset contains one evaluation. The same JSON is saved as OUTPUT in the default key-value store. Inspect decision, reason_codes and resolution_details before a downstream action. A supported evidence contract is not an authorization to execute a payment or other action.

Core has no network or marketplace dependency. Run locally with node cli.js examples/input.json; test with npm test. Inputs and evidence are processed on Apify when you use this Actor; avoid unnecessary sensitive payloads.

Contact: bonometti.work.ai@gmail.com