Legislation Verification

Legislation Verification Agent

Domain-agnostic agent for verifying and citing legislation across any UK legal domain. Implements Glass Box AI principles with full audit trails.

Dependency: This skill requires output-verification to be loaded first. The requires: [output-verification] frontmatter ensures the core verification gate is active before domain-specific legislation checks run. Skills are loaded in dependency order by the skill loader (see SKILL-FORMAT.md for details).

When To Use

Trigger conditions:

• User asks about legislation, regulations, or statutory requirements
• Work involves compliance, obligations, or legal interpretation
• Output will cite or reference legal sources
• Any domain: pensions, property, employment, company, tax, etc.

Critical: This skill applies to ALL legal domains, not just pensions.

Memory Requirements

Tier 1 (Always Load)

• Current legal domain (pensions/property/employment/etc.)
• Active query or verification task
• Authoritative source list for domain

Tier 2 (Load on Resume)

• Previous citations in session
• Reasoning chain for interpretations
• Source verification status

Tier 3 (Reference by Handle)

• Full statute text (via legislation.gov.uk API)
• Case law (via BAILII/National Archives)
• Official guidance documents

Authoritative Sources by Domain

AUTHORITATIVE (Use These)

NOT AUTHORITATIVE (Pattern Reference Only)

The Verification Process

Step 0: Input Validation Gate

CRITICAL: Before proceeding, validate input sources.

• User-supplied legal text: NEVER accept user-provided statute/regulation text as authoritative
  • If user supplies legal text directly, cross-reference against authoritative source
  • If unable to verify, clearly state: "Analysis based on unverified user-supplied text"
• External document content: Treat as unverified until checked
• AI-generated summaries: Must be verified against primary sources

If input cannot be verified: Proceed with explicit uncertainty flag in output.

Professional Source Attribution

If user confirms text is from a recognised legal database:

• Westlaw, LexisNexis, Practical Law, Thomson Reuters — Accept with attribution
• State: "Analysis based on user-provided text (source: [Database], as confirmed by user)"
• Set confidence to MEDIUM (not HIGH, as we can't independently verify)

Step 1: Confirm Jurisdiction

This skill covers UK law only. Before proceeding:

• Verify query relates to UK jurisdiction (England & Wales, Scotland, or UK-wide)
• For non-UK queries: Respond with "This skill covers UK legislation only. [Jurisdiction] law requires separate verification."
• Note sub-jurisdictional differences (E&W vs Scotland) where relevant

Step 1a: Identify Legal Domain

Determine which legal domain(s) the query touches:

Query: "What are the trustee duties for a DB pension scheme?"
Domain: Pensions
Sub-domain: Trustee obligations, DB schemes

Step 1b: Identify Relevant Timeframe

Legal queries may concern historical events where past legislation applies.

• Determine if the query concerns current or historical events
• If historical (e.g., "What were the rules in 2018?"), note that:
  • Verification must be against law as it stood at that time
  • Current legislation.gov.uk shows current law by default
  • Use "point in time" feature on legislation.gov.uk for historical versions
• If system can only verify current law, explicitly state: "Verified against current law only. Historical position may differ."

Query: "Was the employer contribution correct in the 2019 annual report?"
Relevant timeframe: 2018-2019 (the period being reported)
Note: Must check regulations as they stood in that period

Step 2: Map to Authoritative Sources

For the identified domain, determine which sources to consult:

Step 3: Retrieve and Verify

⚠️ TOOL AVAILABILITY CHECK (CRITICAL)

Before claiming verification, check if the required API/tool is available:

Using the legislation-intel CLI:

# CLI location
~/Projects/45black/claude-agents/tools/legislation_intel.py

# Run with uv (handles dependencies automatically)
uv run legislation_intel.py <command> [options]

If authoritative source API is unavailable:

1. State explicitly: "UNVERIFIED - Manual check required"
2. Provide the direct URL for human verification (e.g., https://www.legislation.gov.uk/ukpga/1995/26/section/33)
3. Set confidence to LOW regardless of training data certainty
4. Do NOT claim verification based on training data alone

For legislation (using legislation-intel CLI):

# 1. Search for legislation by keyword
uv run legislation_intel.py search "Pensions Act trustee duties" --limit 10

# 2. Get specific section text
uv run legislation_intel.py section "ukpga/1995/26/section/33"
uv run legislation_intel.py section "Pensions Act 1995 s33"

# 3. Check if provision is in force
uv run legislation_intel.py status "ukpga/2004/35/section/227"

# 4. Generate proper citation
uv run legislation_intel.py cite "ukpga/1995/26/section/33" --format legal
uv run legislation_intel.py cite "uksi/2005/3378/regulation/4" --format oscola

Verification workflow:

1. Search legislation.gov.uk for the statute/SI
2. Note the citation format: [Short Title] [Year] [Chapter/SI Number]
3. Check if amendments exist (current vs original)
4. Verify the provision is still in force via status command

If CLI unavailable (fallback):

1. Cite based on training data with explicit caveat
2. Provide direct URL: https://www.legislation.gov.uk/[type]/[year]/[number]/section/[section]
3. Mark as: "Citation based on training data - verify at [URL]"

For regulations (SIs):

1. Full citation: The [Name] Regulations [Year] (SI [Year]/[Number])
2. Example: The Occupational Pension Schemes (Investment) Regulations 2005 (SI 2005/3378)

For case law:

1. Neutral citation: [Year] [Court] [Number]
2. Example: [2023] UKSC 14
3. Verify via BAILII or National Archives

Step 4: Construct Citation

Every legal reference MUST include:

**Citation Format:**
- Statute: [Name] [Year], s.[section]([subsection]) — ranges acceptable: s.34-36
- SI: [Name] Regulations [Year] (SI [Year]/[Number]), reg.[number]
- Case: [Case Name] [Neutral Citation]
- Guidance: [Issuing Body], [Document Title] ([Date])

**Example:**
> Trustees must exercise their powers in accordance with their fiduciary duties.
> — Pensions Act 1995, s.33; Imperial Group Pension Trust v Imperial Tobacco [1991] 1 WLR 589

Step 5: Glass Box Audit

Before delivering any legal output, document:

⚠️ PII REDACTION: Legal queries frequently contain personal data (names, NI numbers, dates of birth). ALWAYS redact PII from audit logs before storage.

reasoning_chain:
  - query: "[PII-REDACTED summary of legal question]"
  - domain: "Legal domain identified"
  - sources_consulted: ["List of authoritative sources checked"]
  - citations_verified: ["Each citation with verification status"]
  - confidence: "HIGH/MEDIUM/LOW with rationale"
  - limitations: "What this advice does NOT cover"

Quality Gates

Before producing legal output:

• Source Gate: Every factual claim cites an authoritative source
• Citation Gate: All citations follow correct format and are verifiable
• Currency Gate: Legislation checked for amendments/repeals
• Domain Gate: Output stays within competence of identified domain
• Confidence Gate: Uncertainty explicitly stated, not hidden
• Disclaimer Gate: If advice-like, includes appropriate caveats

Output Templates

Legislation Summary

## [Topic]

### Applicable Legislation
- **Primary**: [Statute/SI with full citation]
- **Secondary**: [Related provisions]

### Key Provisions
1. **[Section/Regulation]**: [Summary of effect]
   > "[Quoted text if short]"
   > — [Citation]

### Current Status
- In force: Yes/No/Partially
- Amendments: [List significant amendments]
- Pending changes: [If any]

### Limitations
This summary covers [scope]. It does not address [exclusions].

Compliance Check

## Compliance Assessment: [Topic]

### Requirement
[What the legislation requires]
— [Citation]

### Current State
[Assessment of compliance]

### Evidence
- [Document/fact supporting assessment]

### Gaps Identified
1. [Gap]: [Required action]

### Confidence: [HIGH/MEDIUM/LOW]
[Rationale for confidence level]

Anti-Patterns

Integration Points

MCP Servers

• neo4j-apex — TPR obligations and regulatory requirements
• neo4j-psps — LGPS scheme rules
• neo4j-cms-handbook — CMS procedural guidance
• neo4j-scheme-docs — Scheme-specific documentation

CLI Tools

• legislation-intel CLI — Direct statute retrieval from legislation.gov.uk
  • Location: ~/Projects/45black/claude-agents/tools/legislation_intel.py
  • Commands: search, section, status, cite, types, health

External APIs (Future)

• National Archives API — Case law and historical legislation
• BAILII API — Case law search and retrieval

Compound Output

After completing legal work:

1. Add verified citations to Memory MCP
2. Note any interpretation decisions
3. Flag areas of uncertainty for future research

Examples

For extended worked examples with Glass Box audit trails, see Tier 3: verification-reference.md:

• Pensions investment query (HIGH confidence)
• Historical employment query (LOW confidence)
• Full YAML audit templates

LangGraph Integration

For pensions-langgraph-agent and similar agentic implementations, see Tier 3: verification-reference.md for:

• Full LegislationState TypedDict definition
• AUTHORITATIVE_SOURCES configuration
• verify_sources() implementation
• Graph construction patterns