Run the Chain-of-Verification (CoVe) — procedure

Step-by-step procedure for running Chain-of-Verification before a final answer under an explicit evidence boundary.

Use this procedure to verify claims before final output

Choose this procedure when the answer needs a structured verification loop before the final response instead of a single-pass output.

Core method
Chain-of-Verification before final release
This procedure runs a structured verification loop before emitting the final answer.
Verification comes before delivery, not after.
Required boundary
You must select an evidence boundary first
CoVe does not define admissible evidence on its own.
Choose artifacts-only, authoritative sources, or web verification first.
Fail-closed rule
Exact sentinels override CoVe output sections
If the active evidence boundary requires failing closed, output the exact boundary sentinel and stop even if CoVe is installed.
No partial CoVe output when the boundary fails closed.

Canonical links

Open the linked policy, prompt file, and adjacent references that implement this verification workflow.

Related
Policy and prompt files for this guide

Open the policy linked to this guide and any prompt files used with it.

Verify claims before final output — policy
Prompt files (8) Hide prompt files

Choose a mode

Pick one evidence mode first, then run CoVe inside that boundary.

Option 1 — Artifacts-only CoVe
Verify claims using only the artifacts you provide and no external sources.
Jump to Option 1 →
Option 2 — Authoritative sources CoVe
Verify world-claims using authoritative sources with stable locators.
Jump to Option 2 →
Option 3 — Web verification CoVe
Use browsing or search, if available, plus inline citations and a Sources list.
Jump to Option 3 →

Set up the procedure

Complete these steps before you run the verification loop.

Step 2
Install the boundary system prompt
Install the boundary’s system prompt template for the mode you selected.
Step 6
Run CoVe
Write DRAFT, derive checkable VERIFICATION_QUESTIONS for each factual claim, answer independently in VERIFICATION_ANSWERS using admissible evidence only, then emit FINAL with only verified claims.

Output contract

The CoVe output uses exact sections in order unless the active boundary fails closed first.

DRAFT
Initial answer draft before claim verification.
VERIFICATION_QUESTIONS
Checkable questions derived from the draft’s factual claims.
VERIFICATION_ANSWERS
Independent answers using only admissible evidence.
FINAL
Final answer containing only verified claims.
Exact fail-closed sentinels
If the active evidence boundary requires failing closed, output the exact sentinel and stop.
"HANDS UP – no artifact, cannot verify." · "HANDS UP – no source, cannot verify." · BROWSING_UNAVAILABLE · INSUFFICIENT_EVIDENCE
Boundary priority
Boundary failure behavior overrides CoVe section output.
Do not emit CoVe sections when the active evidence rule requires sentinel-only failure.

Verify the setup

Use this smoke test to confirm that CoVe and the active boundary are wired correctly.

Option 1 expected
Ask a factual question without providing artifacts.
Expected: output exactly "HANDS UP – no artifact, cannot verify."
Option 2 expected
Ask a factual question without providing authoritative sources.
Expected: output exactly "HANDS UP – no source, cannot verify."
Option 3 expected
Ask a factual question without artifacts or citations in web mode.
Expected: output exactly BROWSING_UNAVAILABLE if browsing/search is unavailable, or INSUFFICIENT_EVIDENCE if evidence is insufficient.

Choose the exact mode

Each mode maps to a different evidence boundary and enforcement stack.

Option 2
Authoritative sources CoVe
Verify world-claims using authoritative sources with stable locators.
Use this when claims must be supported by DOI, standard-id, clause, or official vendor doc version and section.
Open policy
Evidence boundary Boundary prompt CoVe prompt
More optional prompts Hide optional prompts
Confidence score Instruction hierarchy
Example: “What does standard/spec X say about Y?” You must provide DOI, standard-id + clause, or official vendor doc version + section.
Option 3
Web verification CoVe
Use browsing or search, if available, plus inline citations and a Sources list.
Use this when the question depends on current or niche public information and requires controlled browsing.
Open policy
Web verification policy Boundary prompt Web verification prompt
More browsing prompts Hide browsing prompts
CoVe prompt Browsing runner Citations contract Confidence score Instruction hierarchy Request web browsing
Example: “Find the most recent official guidance about X and cite it.” You must provide topic X, constraints, and required recency window.

Common mistakes

These are the main failure modes this procedure is intended to block.

Installing CoVe before choosing a boundary
CoVe must run inside an explicit evidence boundary, not before it.
Expecting CoVe sections after boundary failure
If the active boundary requires sentinel-only failure, CoVe sections must not appear.
Missing citations contract in web mode
Omitting citations-output-contract.user.txt in web mode causes the required Sources list format to fail.
Mixing evidence boundaries in one run
Do not mix artifacts-only and authoritative-source boundaries without an explicit selection.
No recency window for “latest”
In web verification mode, asking for “latest” without a recency window weakens the run contract.

Related libraries

Open the matching library when you need the broader policy set, adjacent procedures, prompt files, or the content map.

Policies
Browse the broader policy library and adjacent evidence-boundary rules.
Open index →
How-to
Browse the procedural library and adjacent verification workflows.
Open index →
Prompt library
Open the prompt files used by the CoVe workflow.
Open library →
Content map
Open the broader site structure and adjacent pages.
Open page →