Implement pipeline decomposition (Opus / Sonnet / Haiku stages)

[cpitzi]

Summary

Decompose the monolithic single-Opus-session auditor into a multi-stage pipeline with model-appropriate task assignment.

See docs/architecture.md for the full design document.

Stages

1. Ingestion + Parsing (Haiku) — Parse references into structured JSON
2. Procedural Verification (Sonnet + web search) — DOI resolution, PubMed queries, retraction checks
3. Forensic Interpretation (Opus) — Heuristic application, risk classification, adversarial reasoning
4. Report Generation (Haiku) — HTML report assembly

Key Design Decisions Needed

• Orchestration layer: Simple Python script vs. API tool use vs. workflow engine?
• Schema contracts: Define JSON schemas for each stage boundary
• Parallelization: Per-reference in Stage 2, or batched?
• Error handling: Partial Stage 2 failures — how does Stage 3 handle incomplete data?

Implementation Phases

1. Define and test JSON schemas for inter-stage data
2. Implement Stage 1 (Haiku parsing) as standalone prompt
3. Implement Stage 2 (Sonnet verification) as standalone prompt
4. Implement Stage 3 (Opus forensics) consuming Stage 1+2 output
5. Implement Stage 4 (Haiku report generation) consuming all prior stages
6. Build orchestration script to chain stages
7. Benchmark cost-per-run vs. monolithic v3

Acceptance Criteria

• Each stage independently testable with defined input/output schemas
• Full pipeline produces equivalent output quality to monolithic v3
• Cost-per-run reduced by ≥40% vs. monolithic Opus
• Processing time comparable or better (parallelization in Stage 2 should help)
• Graceful degradation when Stage 2 verification is incomplete

Dependencies

• v3 prompt stable (baseline for comparison)
• Test sets committed for regression testing