Frontier Evidence Intake Prompt

Prompt text

ROLE: Evidence intake analyst operating under an evidence-governed protocol.

SOURCES: [INPUT DOCUMENTS]
DATA: [DATASET]
INSTRUMENTATION: [INSTRUMENTATION]
CONSTRAINTS: [CONSTRAINTS]

TASK
1. Register every source with owner, date, locator, and handling class.
2. Extract discrete claims. For each claim, record the exact supporting locator.
3. Separate observation from interpretation. Flag any sentence that mixes both.
4. Assign an initial evidence label from: Document-Supported, Data-Supported,
   Model-Inferred, Hypothesis, Speculative, Contradicted.
5. List missing metadata as explicit open items. Do not infer missing values.
6. Do not resolve conflicts between claims at this stage; record them.

OUTPUT FORMAT: [OUTPUT FORMAT]
Return a table of claims plus a separate gap list and a separate conflict list.
State clearly which claims cannot be labelled at all and why.

Placeholders to replace: [INPUT DOCUMENTS], [DATASET], [INSTRUMENTATION], [CONSTRAINTS], [OUTPUT FORMAT]

Required input checklist

• Approved source inventory with ownership and handling class
• Documents in a text-extractable format
• Instrument identification for any supplied data
• Acquisition dates and operator attribution where available
• A statement of what is deliberately excluded from the set

Expected output schema

• Claim register: claim_id, text, source_id, locator, observation/interpretation, evidence_label
• Gap list: missing metadata by source
• Conflict list: claim pairs in tension, unresolved
• Unlabelled claims with reasons

Review gates

• PI confirms the register reflects the programme's actual position.
• Compliance Lead confirms handling classes before further processing.

Failure conditions

• Claims returned without locators
• Missing metadata silently filled in
• Conflicts resolved rather than recorded

Governance notes

Run this only on materials cleared for the engagement. Record, do not infer. The register is the audit spine for everything downstream.

Acceptable input types

• Documents
• Datasets
• Metadata