Hallucination Red-Team

Purpose

Force a draft legal output to account for every factual claim, legal claim, cited authority, user-provided fact, model-generated statement, unsupported claim, and uncertainty marker. The skill produces either a safer revised version or a "not ready for use" finding.

This focused hallucination pass complements the broader Red-Team Verifier. It does not independently verify the law and does not certify that a draft is correct.

Use When

• A draft contains legal claims, factual claims, citations, quotations, or confident conclusions that may be unsupported.
• A user asks for a hallucination check, unsupported-claim audit, or safer revised version.
• A draft will be sent externally, filed, used in a client communication, or relied on for a high-risk decision.
• A model produced the draft without a clear claim/source table.

Required Inputs

• The complete draft to red-team.
• All user-provided facts and source documents available for comparison.
• Any cited authorities, research notes, or source lists.
• The intended use and recipient.

If the draft is missing, stop and request it. If sources are missing, continue only as a limitation-heavy hallucination risk screen.

Do Not Use When

• The user wants new legal analysis instead of a verification pass.
• The output has no claims to audit.
• The user asks for a final certification that no hallucinations exist.

Legal Safety Rules

• Produce draft legal work product for attorney review. This is not legal advice.
• Do not claim the absence of hallucinations. State only what was checked and what remains unresolved.
• Do not invent sources, facts, citations, or safer substitute law.
• Do not assert that an authority exists or is fabricated unless verified by source material available in the session.
• Preserve privilege and confidentiality.

Workflow

1. Segment the draft. Break the draft into discrete factual claims, legal claims, citations, quotations, assumptions, recommendations, and conclusions.
2. Classify information origin. For each item, mark whether it is user-provided, source-provided, model-generated, inferred, or unclear.
3. Inventory authorities. List every authority, citation, quotation, source URL, statute, rule, case, regulation, or guidance reference.
4. Flag unsupported claims. Mark claims with no source, insufficient source support, contradicted source support, unclear origin, or overbroad phrasing.
5. Flag authority risks. Mark invented-looking or unverifiable authorities, missing pin cites, jurisdiction mismatch, outdated authority risk, and unsupported legal propositions.
6. Mark uncertainty. Add or recommend [CONFIRM: ...], [VERIFY: ...], [citation needed], [pin cite needed], [verify jurisdiction], and [deadline verification required] markers where needed.
7. Decide readiness. If material unsupported claims, authority risks, or missing gates remain, mark "not ready for use." Otherwise mark "ready for attorney review," not final use.
8. Revise safely. Produce a safer revised version only where the revision removes overclaiming, labels uncertainty, or narrows to supported facts. Do not supply missing law.

Output Format

Deliver:

1. Draft Label — "Draft legal work product for attorney review. Not legal advice."
2. Readiness Finding — Ready for attorney review / Not ready for use, with reasons.
3. Claim-Origin Table

1. Authority Risk Table

1. Unsupported Claims List
2. Uncertainty Markers To Add
3. Safer Revised Version — Or "No safe revision without additional sources/attorney judgment."
4. Attorney Verification Checklist

Attorney Verification Checklist

• [ ] Every factual claim is traced to user-provided or source-provided material, or is marked as unsupported.
• [ ] Every legal claim is supported by verified authority or flagged for attorney research.
• [ ] Every cited authority and quotation has been independently checked.
• [ ] Model-generated claims have been removed, sourced, or marked for verification.
• [ ] Jurisdiction and deadline gaps are resolved or visibly flagged.
• [ ] The revised draft does not present unsupported analysis as settled law.
• [ ] The output is ready for attorney review only; it has not been treated as final.

Full raw SKILL.md

---
name: Hallucination Red-Team
description: "Use when stress-testing draft legal work product for hallucination risk by separating factual claims, legal claims, cited authorities, user-provided information, model-generated information, unsupported claims, uncertainty, and safer revisions."
practice_area: legal-methodology
task_type: verification
jurisdictions: []
risk_level: high
requires_attorney_review: true
inputs:
  - "The complete draft legal output to red-team"
  - "The user-provided facts, source documents, and cited authorities"
  - "The intended use and audience of the output"
outputs:
  - "Hallucination risk report"
  - "Safer revised version or not-ready finding"
related_skills:
  - skills/legal-methodology/red-team-verifier/SKILL.md
  - skills/legal-methodology/citation-integrity-check/SKILL.md
  - skills/legal-methodology/source-validation/SKILL.md
tags:
  - legal-methodology
  - hallucination-check
  - red-team
  - unsupported-claims
  - verification
---

# Hallucination Red-Team

## Purpose

Force a draft legal output to account for every factual claim, legal claim, cited authority, user-provided fact, model-generated statement, unsupported claim, and uncertainty marker. The skill produces either a safer revised version or a "not ready for use" finding.

This focused hallucination pass complements the broader Red-Team Verifier. It does not independently verify the law and does not certify that a draft is correct.

## Use When

- A draft contains legal claims, factual claims, citations, quotations, or confident conclusions that may be unsupported.
- A user asks for a hallucination check, unsupported-claim audit, or safer revised version.
- A draft will be sent externally, filed, used in a client communication, or relied on for a high-risk decision.
- A model produced the draft without a clear claim/source table.

## Required Inputs

- The complete draft to red-team.
- All user-provided facts and source documents available for comparison.
- Any cited authorities, research notes, or source lists.
- The intended use and recipient.

If the draft is missing, stop and request it. If sources are missing, continue only as a limitation-heavy hallucination risk screen.

## Do Not Use When

- The user wants new legal analysis instead of a verification pass.
- The output has no claims to audit.
- The user asks for a final certification that no hallucinations exist.

## Legal Safety Rules

- Produce draft legal work product for attorney review. This is not legal advice.
- Do not claim the absence of hallucinations. State only what was checked and what remains unresolved.
- Do not invent sources, facts, citations, or safer substitute law.
- Do not assert that an authority exists or is fabricated unless verified by source material available in the session.
- Preserve privilege and confidentiality.

## Workflow

1. **Segment the draft.** Break the draft into discrete factual claims, legal claims, citations, quotations, assumptions, recommendations, and conclusions.
2. **Classify information origin.** For each item, mark whether it is user-provided, source-provided, model-generated, inferred, or unclear.
3. **Inventory authorities.** List every authority, citation, quotation, source URL, statute, rule, case, regulation, or guidance reference.
4. **Flag unsupported claims.** Mark claims with no source, insufficient source support, contradicted source support, unclear origin, or overbroad phrasing.
5. **Flag authority risks.** Mark invented-looking or unverifiable authorities, missing pin cites, jurisdiction mismatch, outdated authority risk, and unsupported legal propositions.
6. **Mark uncertainty.** Add or recommend `[CONFIRM: ...]`, `[VERIFY: ...]`, `[citation needed]`, `[pin cite needed]`, `[verify jurisdiction]`, and `[deadline verification required]` markers where needed.
7. **Decide readiness.** If material unsupported claims, authority risks, or missing gates remain, mark "not ready for use." Otherwise mark "ready for attorney review," not final use.
8. **Revise safely.** Produce a safer revised version only where the revision removes overclaiming, labels uncertainty, or narrows to supported facts. Do not supply missing law.

## Output Format

Deliver:

1. **Draft Label** — "Draft legal work product for attorney review. Not legal advice."
2. **Readiness Finding** — Ready for attorney review / Not ready for use, with reasons.
3. **Claim-Origin Table**

   | # | Claim | Type | Origin | Support status | Risk | Required action |
   |---|---|---|---|---|---|---|

4. **Authority Risk Table**

   | Authority | Use in draft | Verification status | Hallucination risk | Required action |
   |---|---|---|---|---|

5. **Unsupported Claims List**
6. **Uncertainty Markers To Add**
7. **Safer Revised Version** — Or "No safe revision without additional sources/attorney judgment."
8. **Attorney Verification Checklist**

## Attorney Verification Checklist

- [ ] Every factual claim is traced to user-provided or source-provided material, or is marked as unsupported.
- [ ] Every legal claim is supported by verified authority or flagged for attorney research.
- [ ] Every cited authority and quotation has been independently checked.
- [ ] Model-generated claims have been removed, sourced, or marked for verification.
- [ ] Jurisdiction and deadline gaps are resolved or visibly flagged.
- [ ] The revised draft does not present unsupported analysis as settled law.
- [ ] The output is ready for attorney review only; it has not been treated as final.