Section Validator

SKILL: section-validator

Purpose

Build and maintain the 4 deterministic verification gates (Stage 3). Gates run on the full draft text with zero LLM calls. They validate, auto-fix formatting, and flag issues for review.

When to Use

• Building or modifying any gate
• Adding new entity extraction patterns
• Debugging false positives / false negatives
• Extending verified provisions coverage

Architecture Context (v5.1 — what's running)

4 gates run sequentially on draft.draft_artifacts[0].text (full document text). Total time: ~0.1s.

draft_freetext -> evidence_anchoring -> lkb_compliance -> postprocess -> citation_validator -> review

All gates are deterministic — zero LLM calls.

Gate 1: Evidence Anchoring (evidence_anchoring.py)

For each extracted entity (date, amount, reference) in the draft:

1. Check if exists in intake facts, evidence, or user_request
2. If YES -> keep as-is
3. If NO -> replace token with {{PLACEHOLDER}}

Entity Parsers

• Date parser: handles 15.03.2024, 15/03/2024, 15th March 2024, etc. Excludes Section/Article/Order numbers
• Amount parser: handles Rs. 20,00,000/-, INR 20 lakh, Rupees Twenty Lakhs. Indian comma notation. Excludes years/days/percentages
• Reference parser: cheque no., UTR, NEFT Ref — keyword-aware

Output

• evidence_anchoring_issues — list of dicts with type and description
• Modified draft text with unsupported tokens replaced

Gate 2: LKB Compliance (lkb_compliance.py)

Checks draft against LKB entry for the classified cause type:

1. Primary acts cited — are the LKB primary_acts referenced in draft?
2. Superseded law check — auto-replaces old acts with current (e.g., IPC -> BNS, CrPC -> BNSS)

Auto-Fix

• Superseded law references auto-replaced (deterministic substitution)
• Missing primary acts -> flagged as issue for review

Output

• Issues merged into postprocess_issues

Gate 3: Postprocess (postprocess.py)

Deterministic formatting fixes on the draft text:

1. Paragraph numbering — fix sequential numbering gaps
2. Annexure labels — ensure Annexure A, B, C format
3. Verification clause — ensure Order VI Rule 15 CPC clause present
4. Spacing — normalize whitespace, line breaks
5. and/or fix — replace "and/or" with "and" (Indian legal convention)
6. Paragraph breaks — split run-together numbered paragraphs
7. Annexure list — split semicolon-separated items
8. Prayer items — ensure prayer sub-items on separate lines

Output

• postprocess_issues — list of formatting fixes applied
• Modified draft text with all fixes applied

Gate 4: Citation Validator (citation_validator.py)

Every statutory citation in the draft must:

1. Exist in verified_provisions (from enrichment) OR be a well-known procedural provision
2. Not be a fabricated case citation (AIR/SCC/ILR)

Case Citation Patterns (flagged as ERROR)

• (2015) 5 SCC 123
• AIR 2015 SC 123
• 2015 SCC OnLine SC 123
• ILR 2015 KAR 123

Well-Known Provisions (always allowed)

• CPC: Section 26, Order VI Rule 15, Order VII Rule 1, Section 34, Section 151
• Constitution: Article 14, 19, 21, 226, 32
• Evidence Act: Section 65B

Output

• citation_issues — list of dicts with severity and message
• Enabled/disabled via DRAFTING_CITATION_VALIDATOR_ENABLED setting

Key Files

Settings

DRAFTING_CITATION_VALIDATOR_ENABLED: bool = True

Rules

• NEVER modify legal substance — only replace unsupported tokens or flag claims
• NEVER remove sentences — only flag them
• Entity extraction must be deterministic (no LLM)
• Gates run on full document text (draft.draft_artifacts[0].text)
• False positives are better than false negatives
• All gates run regardless of individual failures (graceful degradation)
• Gate issues are summarized for review via _build_gate_errors_summary()

Anti-Patterns

• Do NOT use LLM for any gate — deterministic only
• Do NOT block pipeline on gate warnings — flag and continue
• Do NOT auto-remove flagged citations — let review handle
• Do NOT flag placeholder usage as an issue
• Do NOT auto-fix substantive legal content — only formatting