v0.1 draft · core spec

Citations that travel
with the claim.

AI Evidence Format is an open JSON spec for structured citations attached to LLM-generated claims: source URI + type + publisher, span (text-quote / CSS selector / page range), retrieval method + confidence + freshness, content-hash verification, and synthesis role (supporting / contradicting / partial / background).

  • Detect via evidence_version
  • Well-known: — (travels with the LLM response)
  • Cross-referenced from AI Incident Cards on failure

Why structured citations?

Hallucinations get a lot of airtime; untraceable citations get less. An AI tool can cite a real-sounding URL that no longer resolves, or quote text that doesn't appear at the cited location, or paraphrase so loosely that no fact-checker can find the source. AI Evidence Format is the structured artifact that makes each citation independently verifiable.

Claim text

The literal claim being supported. A grader can read it standalone before chasing the citation.

Source

uri, type (document / webpage / api / book / paper), title, publisher, fetched_at. The minimum auditable identification of where the claim came from.

Span

selector_type{text_quote, css_selector, fragment_identifier, page_range} with selector_value and optional exact_text. Points at the bytes that support the claim.

Retrieval

method (semantic / keyword / hybrid / direct), confidence, rank, freshness_age_seconds. Disclosure of how the source was found.

Verification

content_hash — canonical SHA-256 over the cited bytes. Optional signature. A reviewer can recompute the hash and prove the cited content didn't change after retrieval.

Synthesis role

supporting / contradicting / partial / background. Forces the agent to declare how the citation relates to the claim.

The required sections

  1. evidence_version — must be "0.1"
  2. evidence_id — stable identifier
  3. claim_text — the literal claim being supported
  4. source — URI + type + title + publisher + fetched_at
  5. span — selector_type + value + exact_text
  6. retrieval — method + confidence + rank + freshness
  7. verification — canonical content_hash, optional signature
  8. synthesis_role — supporting / contradicting / partial / background

A canonical example

{
  "evidence_version": "0.1",
  "evidence_id": "ev-2026-05-12-a4f9c1",
  "claim_text": "The AEO Protocol defines three pillars: Declare, Discover, and Audit.",
  "source": {
    "uri": "https://github.com/mizcausevic-dev/aeo-protocol-spec/blob/main/SPEC.md",
    "type": "document",
    "title": "AEO Protocol v0.1 — Specification",
    "publisher": "mizcausevic-dev",
    "fetched_at": "2026-05-12T03:45:00Z"
  },
  "span": {
    "selector_type": "text_quote",
    "selector_value": "## 3. The three pillars",
    "exact_text": "## 3. The three pillars\n\n### 3.1 Declare\n\nThe entity publishes a single JSON document at a fixed path describing itself."
  },
  "retrieval": { "method": "hybrid", "confidence": 0.93, "rank": 1, "freshness_age_seconds": 90 },
  "verification": { "content_hash": "sha256:c7d1f2a3b4c5d6e7f8a9b0c1d2e3f4a5b6c7d8e9f0a1b2c3d4e5f6a7b8c9d0e1" },
  "synthesis_role": "supporting"
}

About the Kinetic Gain Protocol Suite

AI Evidence Format is one of ten open JSON specifications in the Kinetic Gain Protocol Suite. Five core specs plus the EdTech trio, the HealthTech extension, and the cross-cutting Incident Card. Front door: suite.kineticgain.com.