Evidence clearing protocol

Digest-bound guide

Conflict and uncertainty for machine-generated claims

How should conflict and uncertainty be recorded for machine-generated claim?

A fail-closed protocol for conflict and uncertainty when evaluating machine-generated claims.

Bounded answer

Record conflicting evidence as separate, versioned propositions with their scopes and uncertainty; do not average disagreement into a confident sentence or choose a preferred source silently. For machine-generated claims, preserve the generated claim, model and prompt commitments, cited source, locator, and human review state.

Input contract

What must be fixed first

Claim and scope

Each conflicting source and locator

Version and publication relationship

Declared uncertainty or unresolved question

Procedure

Work the decision in order

  1. 1

    Normalize the propositions without erasing scope.

  2. 2

    Inspect each cited passage independently.

  3. 3

    Classify direct contradiction, scope difference, or version change.

  4. 4

    Preserve uncertainty and missing evidence.

  5. 5

    State what evidence could resolve the conflict.

Expected outputs

  • Typed conflict record
  • Scope comparison
  • Unresolved-evidence requirement

Refuse when

  • One side has not been inspected.
  • Different populations or definitions are called contradictions.
  • A later source silently overwrites historical evidence.
  • A model citation is a lead until the cited passage is inspected and shown to carry the claim.

Questions this guide answers

How should conflict and uncertainty be recorded for machine-generated claim?

Record conflicting evidence as separate, versioned propositions with their scopes and uncertainty; do not average disagreement into a confident sentence or choose a preferred source silently. Preserve the generated claim, model and prompt commitments, cited source, locator, and human review state.

What must be preserved for machine-generated claims?

The generated claim, model and prompt commitments, cited source, locator, and human review state.

When must the protocol refuse?

It must refuse when one side has not been inspected. or when the identified source cannot be inspected at the required depth.

Does a successful preflight verify the claim?

No. It shows that the evidence request is sufficiently structured to begin inspection; factual verification remains a separate stage.

What does this protocol not establish about machine-generated claims?

A model citation is a lead until the cited passage is inspected and shown to carry the claim.

Limits

  • A model citation is a lead until the cited passage is inspected and shown to carry the claim.
  • This protocol produces a preflight or evidence-structure finding, not a verified Evidence Dossier.
  • No source is treated as inspected merely because its metadata or identifier resolves.