V9 Architecture

SKILL: v9-architecture

Use when: planning, building, or reviewing v11.0 architecture components (LKB 2-layer model, document schemas, structured prompt builder, gates, family migrations).

v11.0 Architecture — Scalable Context-Driven Pipeline

Core Principles

1. Better context to LLM = better draft — no complex engine needed
2. Separate law from structure — cause type (92) × document type (12) = 1,104 combinations
3. Decide law before drafting, enforce law after drafting

The Root Problem

_build_lkb_brief_context() dumps 13 instruction categories as flat text at equal priority (~400 lines). LLM ignores half, hallucinates the rest. This is why drafts have thin facts, missing sections, wrong citations.

The Solution: LKB 2-Layer + Document Schemas → Structured Prompt

LKB Layer 1 (legal knowledge)  +  LKB Layer 2 (document components)  +  Document Schema
            ↓                              ↓                                    ↓
         COMBINE into ONE structured prompt with clear hierarchy
            ↓
         LLM drafts complete document (1 call)
            ↓
         Gates validate (0 LLM)

LKB 2-Layer Model

Document Schemas (12 schemas)

Trial court:  plaint, written_statement, rejoinder, counter_claim
Applications: interim_application, condonation_of_delay, set_aside_ex_parte, caveat
Appellate:    appeal_memo, revision_petition, review_petition
Post-decree:  execution_application

Each schema defines: section order, per-section instructions, filed_by, annexure_prefix, verification_type, filing_rules.

Structured Prompt (replaces flat dump)

═══ DOCUMENT STRUCTURE (follow this section order exactly) ═══
[from document schema]

═══ LEGAL DATA (cite ONLY from this) ═══
[from LKB Layer 1 + Layer 2]

═══ FACTS GUIDANCE ═══
[from LKB facts_must_cover + red_flags]

═══ CLIENT FACTS ═══
[from user intake]

~1,500 tokens. Clear hierarchy. 3 sections, not 13.

Pipeline (v11.0)

intake_classify (cause_type + document_type)
  → enrichment (LKB + document schema)
  → draft (structured prompt → LLM → complete document)
  → gates (evidence_anchoring, lkb_compliance, postprocess, citation_validator)
  → review (optional)
  → END

Scaling

New cause type:     1 LKB entry → works with ALL 12 doc types
New document type:  1 schema → works with ALL 92 cause types
New domain:         N LKB entries + reuse schemas → 0 code changes

What Stays From v5.1/v9.0

• All verification gates (7 structural + semantic, zero LLM)
• All applicability compilers (5 families)
• All consistency gates (7 families, 50+ checks)
• Pipeline structure (intake → enrich → draft → gates → review)
• 2 LLM calls (intake + draft), review optional

Migration Path

1. Enrich LKB with Layer 2 (available_reliefs with prayer_text, jurisdiction_basis)
2. Create 3 schemas (plaint, written_statement, interim_application)
3. Rewrite prompt builder (~100 lines replaces 400-line _build_lkb_brief_context)
4. Add document_type to intake classification
5. Add remaining 9 schemas
6. Remove old engine.py (-2400 lines)

Civil Plugin — Families (unchanged)

Key Files

• Draft prompt (to rewrite): nodes/draft_single_call.py (_build_lkb_brief_context)
• Template engine (to remove): templates/engine.py
• LKB: lkb/causes/ (16 sub-group files, 92 entries)
• LKB helpers: lkb/causes/_helpers.py (_entry() schema)
• Schemas (NEW): schemas/ directory
• Civil decision: nodes/civil_decision.py
• States: states/draftGraph.py

Rules

• Do NOT dump flat text to LLM — use structured prompt with hierarchy
• Do NOT mix law data with document structure — Layer 1 and Layer 2 are separate concerns
• Do NOT build per-cause-type engine builders — LLM drafts from structured prompt
• Do NOT add per-document-type hardcoded logic — schema drives structure
• Do NOT let LLM decide legal track — applicability compiler decides
• Do NOT add LLM calls for deterministic tasks
• LKB data correctness is THE critical requirement — wrong data = wrong draft
• Document schema correctness is critical — wrong structure = court rejects