Run the fact-checking kit — procedure

End-to-end fact-checking workflow using evidence boundaries, web verification, Chain-of-Verification, and confidence scoring.

Purpose

Use this page to run an end-to-end fact-checking workflow that reduces factual errors and improves auditability.

Enforcement (fail-closed):

You must choose an evidence boundary first and follow the active policy’s admissible-evidence rules.
If required evidence is missing, you must fail closed using the exact sentinel required by the active system prompt template:
- "HANDS UP – no artifact, cannot verify."
- "HANDS UP – no source, cannot verify."
- BROWSING_UNAVAILABLE
- INSUFFICIENT_EVIDENCE

Option 1 (Artifacts-only run): verify only from artifacts you provide (no external sources).
Option 2 (Authoritative sources run): verify world-claims using authoritative sources with stable locators.
Option 3 (Web verification run): use browsing/search (if available) + inline citations + Sources list.

1) Choose an evidence boundary and install its policy + system prompt template:

Choose allowed sources for factual answers 2) Install the verification workflow you will use for non-trivial answers:
Chain-of-Verification (CoVe) — procedure 3) If you require a confidence line in successful (non-sentinel) responses, install the confidence score rule:
Add an evidence-based confidence score (0–100) to every response 4) Run your answer generation using CoVe, then enforce claim-level traceability:
Artifacts-only: every factual claim ends with [artifact-id §locator].
Authoritative sources / web verification: every world-claim includes stable locators or inline markers with a Sources list (as required by the active policy).

Ask one factual question without providing any artifacts or citations.

Option 1 expected: output exactly "HANDS UP – no artifact, cannot verify."
Option 2 expected: output exactly "HANDS UP – no source, cannot verify."
Option 3 expected: output exactly BROWSING_UNAVAILABLE (if browsing is unavailable) or INSUFFICIENT_EVIDENCE (if evidence is insufficient).

Example

Question: “Why did this CI run fail?”
You must provide: CI/build logs + referenced config snippets/files. Each factual claim must cite [artifact-id §locator].

Policy (rules): Facts-only: Authoritative sources required
System prompt template (copy/paste): facts-only-authoritative-sources-required.system.txt
Procedure: Facts-only: Authoritative sources required — procedure

Example

Question: “What does standard/spec X say about Y?”
You must provide: DOI or standard-id + section/clause (or official vendor doc version + section). If not verifiable, output exactly "HANDS UP – no source, cannot verify."

Policy (rules): Web Verification & Citations Policy
System prompt template (copy/paste): web-verification-and-citations.system.txt
Procedure: Request web browsing — prompt template

Example

Question: “Find the most recent official guidance about X and cite it.”
You must provide: topic X + constraints (jurisdiction/organization) + required recency window. If browsing is unavailable, output exactly BROWSING_UNAVAILABLE.

Duplicating stack definitions across multiple pages instead of linking to the SSOT procedure.
Mixing evidence boundaries (artifacts-only vs authoritative vs web verification) in the same run without an explicit selection.
Asking for “latest” without specifying a recency window.
Emitting anything other than the exact sentinel when the active policy requires a fail-closed response.