Doublecheck

Doublecheck

Run a three-layer verification pipeline on AI-generated output. Extract every verifiable claim, find sources the user can check independently, flag hallucination patterns, and produce either inline verification or a full report.

When to invoke

• "Turn on doublecheck."
• "Doublecheck that answer."
• "Run full verification on this text."
• "Fact-check these claims with sources."
• "Verify the citations and statistics."

Prerequisites and context

• Use web_search for source verification when claims need external evidence.
• Use assets/verification-report-template.md for full reports.
• Do not claim a statement is true just because a source was found; report the source and let the user decide.

Activation modes

Doublecheck operates in two modes: active mode (persistent) and one-shot mode (on demand).

Classify every active-mode response before sending it.

For inline verification, generate the response normally, add a Verification section, and list each checked claim with a confidence rating and source link where available:

---
**Verification (N claims checked)**

- [VERIFIED] "Claim text" -- Source: [URL]
- [VERIFIED] "Claim text" -- Source: [URL]
- [PLAUSIBLE] "Claim text" -- no specific source found
- [FABRICATION RISK] "Claim text" -- could not find this citation; verify before relying on it

_Say "full report" for detailed three-layer verification with sources._

Auto-escalate to a full report when any claim is DISPUTED or FABRICATION RISK. Place this callout before the report summary:

**Heads up:** I'm not confident about [specific claim]. I couldn't find a supporting source. You should verify this independently before relying on it.

Always use a full report for legal analysis, regulatory interpretation, compliance guidance, case citations, or statutory references.

Layer 1: Self-audit

Re-read the target text with a critical lens. This layer extracts and analyzes claims before any web search.

Layer 2: Source verification

For each extracted claim, search for external evidence and record URLs the user can visit independently.

1. Formulate a search query that would surface a primary source. For citations, search the exact title or case name. For statistics, search the number plus topic. For factual claims, search the key entities and relationship.
2. Run web_search. If the first search misses relevant results, reformulate once with different terms.
3. Decide whether the result directly supports, contradicts, or fails to address the claim.
4. Record the result with the source URL.

Prefer primary and authoritative sources: official documentation, specifications, standards, court records, legislative texts, regulatory filings, peer-reviewed publications, official organizational sites, press releases, and established reference works. Mark news articles, blog posts, and wiki pages as secondary when used.

Citations are the highest-risk category and often look plausible-sounding. Citations get the strictest treatment. Search the exact case name, statute, section number, paper title, standard, or document. If the citation cannot be found at all, rate it FABRICATION RISK.

Layer 3: Adversarial review

Assume the target text contains errors and actively try to find them.

For each major claim, ask: What would make this claim wrong? Is there a common misconception here? Would a subject matter expert object? Is this claim potentially outdated relative to the current date?

Escalate prominently when a citation cannot be found, a statistic has no identifiable source, a legal or regulatory claim contradicts authoritative sources, or a disputed claim is stated with high confidence.

Confidence ratings

Report principles:

• Provide links, not final verdicts.
• Present both sides when sources conflict.
• Mark vague or subjective claims as unfalsifiable when applicable.
• Distinguish "could not verify" from "wrong".
• Lead with the items that need the most attention.

Domain-specific scrutiny

Progressive disclosure and bundled resources

• assets/verification-report-template.md: use for complete three-layer verification reports, auto-escalation reports, and all high-stakes content.

Output template

## Doublecheck verification report

**Status:** verified | needs human review | blocked
**Target:** <text, response, document, or claim set reviewed>
**Claims checked:** <count>

### Summary
| Rating | Count | Notes |
| --- | ---: | --- |
| VERIFIED | <n> | <short note> |
| PLAUSIBLE | <n> | <short note> |
| UNVERIFIED | <n> | <short note> |
| DISPUTED | <n> | <short note> |
| FABRICATION RISK | <n> | <short note> |

### Claim verification
| ID | Claim | Category | Rating | Sources | Notes |
| --- | --- | --- | --- | --- | --- |
| C1 | <claim text> | factual/statistical/citation/entity/causal/temporal | VERIFIED | <URL> | <evidence or caveat> |

### Adversarial review
- <hallucination pattern or red flag checked>: <result>

### Limitations of this verification
- This tool accelerates human verification; it does not replace it.
- Web search results may not include the most recent information or paywalled sources.
- The adversarial review uses the same underlying model that may have produced the original output. It catches many issues but cannot catch all of them.
- A claim rated VERIFIED means a supporting source was found, not that the claim is definitely correct. Sources can be wrong too.
- Claims rated PLAUSIBLE may still be wrong. The absence of contradicting evidence is not proof of accuracy.

Quality gate

• Every verifiable claim has an ID and category.
• Internal contradictions were checked before web search.
• Citations, precise statistics, dates, and high-risk claims were searched with web_search.
• Every source-backed claim includes a URL.
• DISPUTED and FABRICATION RISK claims are escalated before the summary.
• Legal, regulatory, compliance, case-citation, and statutory content uses the full report.
• Full reports follow assets/verification-report-template.md and include limitations.
• Inline verification ends with _Say "full report" for detailed three-layer verification with sources._