AI Research Evidence Verification

AI Research Evidence Verification is a deterministic evidence-quality gate for agents, OSINT pipelines, research workflows, and automated decision systems. It checks whether caller-supplied evidence is structurally sufficient, independent, fresh, and non-contradictory enough to justify a downstream conclusion — without replacing missing information with an LLM guess.

The Actor returns SUPPORTED, UNRESOLVED, or INSUFFICIENT_EVIDENCE, plus the evidence gaps and confidence limits that explain the result.

What it checks

For one structured claim and a set of caller-supplied evidence items, Research Verification evaluates:

• required evidence-class coverage;
• source independence;
• freshness and unknown freshness;
• supporting versus contradicting evidence;
• contradiction ratio;
• a maximum justifiable confidence ceiling;
• source provenance URLs supplied by the caller.

Duplicate evidence from the same source does not become independent corroboration merely because it has another URL or ID. Evidence with unknown freshness is not silently treated as current.

What it does not do

This Actor does not crawl the web, discover evidence, perform forensic capture, establish chain of custody, or determine legal admissibility. It evaluates structured evidence supplied by an upstream retriever, researcher, agent, or application.

Typical uses

Use Research Verification as a pre-publication evidence check, an OSINT evidence-quality gate, a guardrail before an autonomous agent acts on a researched claim, or a machine-readable step between retrieval and decision making.

It is especially useful when a workflow needs to preserve explicit missing evidence instead of forcing a fluent yes/no answer.

Input

Provide a claim, an evidence inventory, and as_of. Evidence items can include a source identity and URL, evidence class, relation to the claim, observation timestamp, and freshness TTL.

{
  "claim": {
    "id": "claim-1",
    "text": "The declared requirement is currently in force.",
    "required_evidence_classes": ["official_source", "independent_source"]
  },
  "evidence": [],
  "as_of": "2026-09-07T12:00:00Z"
}

Output

The default dataset receives one machine-readable report containing the verdict, coverage ratio, missing evidence classes, contradiction IDs and ratio, source-independence ratio, freshness report, confidence ceiling, limiting factor, and provenance URLs.

The confidence ceiling is structural: it limits how confident a downstream system may reasonably be given the declared evidence. It is not a probability that the claim is true.

Pricing

Research Verification uses Apify pay-per-event pricing. One research-verification-report event is charged only when a complete report is written to the default dataset. The beta event price is USD 0.02 per completed report.

Limitations

The result depends on the quality and identity metadata of caller-supplied evidence. Hidden common ownership, copied reporting, inaccurate timestamps, or incorrect upstream extraction cannot be inferred unless represented in the input. Human review remains appropriate for high-impact or ambiguous decisions.

Do not submit passwords, API keys, private keys, personal data, or confidential material that is not necessary for the assessment.

Suez Evidence & Decision Infrastructure for AI Agents

Research Verification is part of the Suez evidence-and-decision tool family. A typical workflow can use Agent Readiness to audit machine-facing publication, Research Verification to test evidence sufficiency, Task Feasibility to preflight execution, Opportunity Decision to make a bounded pursuit decision, and Decision Monitor to compare that decision as evidence or assumptions change. Each Actor remains independently callable.