nicsuzor/osb-drafting

OSB Case Drafting Skill

When to invoke: Preparing analysis for Oversight Board case deliberation.

What it provides: Structured case analysis with IRAC reasoning, position variants, precedent mapping.

What it doesn't do: Make final decisions, substitute for Board deliberation, guess at facts not in record.

Prerequisites

• OSB MCP server running (test: mcp__osb__search available)
• Case announcement or full case document accessible
• If MCP unavailable: HALT and report infrastructure gap (per AXIOMS.md fail-fast)

Workflow

Phase 1: Case Intake

1. Read case announcement - Extract:

   • Content description (format, platform, reach)
   • Procedural history (reports, reviews, appeals)
   • Meta's stated rationale
   • Board's stated selection reason
   • Strategic priorities engaged

2. Identify applicable policies:

   • Primary policy (most direct violation alleged)
   • Secondary policies (related standards)
   • Policy exceptions that might apply
   • Policy gaps exposed by facts

Phase 2: Precedent Research

3. Search OSB jurisprudence using mcp__osb__search:

   • Same/similar policies applied
   • Same content type (AI-generated, crisis, political)
   • Same context (conflict, elections, public health)
   • Same procedural issues (automation, fact-checking)

4. Retrieve key precedents using mcp__osb__get_case_summary:

   • Extract established legal tests
   • Note defined terms and their scope
   • Identify prior recommendations (implemented or not)
   • Document ratio decidendi (binding principles)

Minimum searches per case:

• Primary policy name + "violation"
• Content type + "precedent"
• Context keyword + "crisis" or "conflict"
• Procedural issue (e.g., "automated enforcement" or "fact-checking")

Phase 3: Tension Mapping

5. Identify decision points where reasonable Board members could disagree:

   • Policy interpretation questions
   • Fact characterization questions
   • Balancing/proportionality questions
   • Remedy selection questions

6. Map fault lines for each decision point:

| Dimension | Position A | Position B | | ------------------------------------ | ---------------------- | ------------------------- | | Textualist vs Purposivist | Strict policy text | Harm-based/intent reading | | Speech-protective vs Harm-preventive | Favor expression | Favor safety | | Individual vs Systemic | Case-specific remedy | Policy-level change | | Global vs Contextual | Consistent application | Regional sensitivity |

Phase 4: IRAC Analysis

7. For each major issue, structure as:

### Issue: [Precise legal question]

#### Position A (likely [majority/minority]): [Outcome]

**Rule**: [Applicable policy text + IHRL standard]
**Analysis**: [Application to facts]
**Precedent**: [Case citations with relevance]
**Conclusion**: [Why this position prevails under this reading]

#### Position B (likely [minority/majority]): [Outcome]

**Rule**: [Alternative reading or emphasis]
**Analysis**: [Application highlighting different facts]
**Precedent**: [Contrary or distinguishable cases]
**Conclusion**: [Why this position has merit]

#### What determines the split:

- [Key factual/value difference 1]
- [Key factual/value difference 2]

Phase 5: Recommendations

8. Draft recommendations addressing:
   • Immediate remedy: Content disposition
   • Policy clarification: Terms needing definition
   • Enforcement process: Procedural improvements
   • Transparency: Disclosure requirements

Recommendations must be:

• Actionable (specific enough to implement)
• Measurable (can verify compliance)
• Precedent-consistent (cite supporting prior decisions)

Phase 6: Citation Verification

9. Extract all case citations from draft:

   • Scan for case IDs (format: XX-XXXXXXX or XXX-XXXXXXXX)
   • Note each quoted passage with its attributed source
   • List all claimed principles and recommendations

10. Verify each citation:

For each cited case, always retrieve full document text:

mcp__osb__get_document(record_id, include_full_text=true)

Why full text is required: Summaries are AI-generated abstractions that may not contain exact quoted phrases. Only the full decision text can verify:

• Exact quote accuracy
• Context of statements
• Complete recommendation wording
• Nuanced legal reasoning

| Check | Criteria | Verification Method | | ------------------------- | ---------------------------------------- | ------------------------------------------ | | Case ID exists | Must return valid result | get_document returns content | | Case name matches | Exact or close match | Compare to official title | | Quoted text accurate | Exact phrase in full text | Search for quoted string in document | | Principle attributed | Ratio/reasoning section supports claim | Read legal reasoning | | Recommendation accurate | Matches verbatim recommendation text | Check recommendations section | | Quote originates here | Quote is FROM this case, not cited BY it | Check quote isn't referencing another case |

CRITICAL - Search for each quoted phrase individually:

mcp__osb__search(query="[exact quoted phrase]", n_results=5)

This catches:

• Quotes that don't exist in any case
• Quotes misattributed to wrong case (e.g., Case A cites Case B, draft wrongly attributes to A)
• Paraphrases that distort meaning

Common misattribution pattern: Case A quotes or cites Case B. Draft attributes the quote to Case A. Always check if the quote says "The Board found in [other case]..." or similar - the principle originates from the cited case, not the citing case.

11. Generate verification report:

## Citation Verification Report

### VERIFIED ✓

- [Case ID]: [Cited principle] - CONFIRMED at [location in full text]

### NEEDS CORRECTION ✗

- [Case ID]: [Issue description]
  - Draft says: "[quoted text]"
  - Actual text: "[exact text from full document]"
  - Fix: [correction]

Note: There should be NO "UNVERIFIABLE" category. All citations must be verified against full document text. If OSB MCP is unavailable, HALT per AXIOMS fail-fast principle.

12. Quality gate - Draft NOT complete until:
    • All citations verified against full document text
    • 0 items in NEEDS CORRECTION
    • Corrections made and re-verified

Common error types:

• Case mislabeling (wrong case name for ID)
• Quote fabrication (phrase not in case text)
• Principle misattribution (case doesn't establish cited holding)
• Recommendation misstatement (wrong implementation status)
• Citation chain error (quote is FROM Case B but draft attributes to Case A which merely cites Case B)

Decision Point Checklist

Common tensions to evaluate in every case:

• [ ] Policy scope: Does content fit policy as written?
• [ ] Harm threshold: Standard met (imminent/direct)?
• [ ] Context weight: How much does setting matter?
• [ ] Speaker identity: Newsworthy figure? Official statement?
• [ ] Remedy proportionality: Removal vs label vs nothing?
• [ ] Regional equity: Comparable treatment globally?
• [ ] Automation role: Human review required?
• [ ] Fact-checker capacity: Realistic at scale?

Human Rights Framework

Apply international standards consistently:

• ICCPR Art. 19: Expression (necessity, proportionality, legality)
• ICCPR Art. 25: Political participation
• Rabat Plan: Six factors for incitement analysis
• UNGPs: Business responsibility for human rights

Output Format

Full Case Draft

# [Case Title] - Draft Analysis

## Summary

[2-3 sentence overview]

## Facts

[Key facts from record]

## Procedural History

[Reports, reviews, appeals, Meta's position]

## Applicable Policies

- Primary: [Policy name and text]
- Secondary: [Related policies]

## Issue 1: [First major question]

[Full IRAC with position variants per Phase 4]

## Issue 2: [Second major question]

[Full IRAC with position variants]

[Continue for each issue]

## Recommendations

1. [Recommendation with precedent support]
2. [Recommendation with precedent support]

## Precedent Map

| Case | Relevance | Key Principle |
| ---- | --------- | ------------- |
| [ID] | [Topic]   | [Holding]     |

Quality Checks

Before finalizing draft:

• [ ] Every legal claim has specific precedent citation
• [ ] Both sides of each tension articulated fairly
• [ ] Human rights standards applied consistently
• [ ] Recommendations are actionable and measurable
• [ ] No speculation beyond case record
• [ ] Position variants reflect genuine disagreement potential
• [ ] Phase 6 verification complete: All citations verified, 0 corrections needed

Anti-Patterns

Avoid:

• Guessing at Board member positions (analyze arguments, not people)
• Over-reliance on single precedent (triangulate multiple sources)
• Ignoring contrary precedent (address and distinguish)
• Policy advocacy (present options, don't predetermine outcome)
• Ungrounded factual claims (cite record or mark as unknown)