Back to research
OPEN SOURCE MIT License — 2026

Codex Legal Engine

Multi-model architecture for legal analysis. Externalized calibration, epistemic labeling, active verification.

The Problem

A jurist using LLM for legal analysis faces a structural problem: the model doesn't distinguish established from debated, positive law from minority doctrine. It produces fluent, confident, potentially false responses. Codex is an architectural response.

The Architecture

Four sequential steps:

1. Externalized Calibration

Model: Claude Haiku

6 probes (R1–R3 rigor, P1–P3 scope/depth), 3 cursors. Calibrator never touches legal data.

2. Layer Selection

Epistemic Regimes: Four layers

  • Litigation
  • Advisory
  • Exploratory
  • Monitoring

Each defines what the model may say and with what certainty.

3. Labeled Analysis

Model: Claude Sonnet or Opus

Epistemic Levels: Five labels

  • [ESTABLISHED]
  • [MAJORITY POSITION]
  • [ARGUED CONNECTION]
  • [DEBATED]
  • [PROSPECTIVE]

4. Active Verification

Protocol: CoVe (Consensus Verification)

  • Existence
  • Content
  • Contradiction

Transversality Matrix

17 adjacent domains with 0–3 scoring.

Direct Access to Case Law

7 MCP tools connected to Judilibre API (Cour de cassation):

  • 4 search engines
  • Full text access
  • Pourvoi verification
  • Taxonomy lookup

Seven Posture Rules

  1. Presumption of expertise
  2. Systematic gap signaling
  3. Maximum informational density
  4. Positional neutrality
  5. Anti-acquiescence
  6. Traceability
  7. Stylistic sobriety

Installation and Cost

Claude-as-installer pattern:

  • Claude Pro: €18/month
  • Make: €9/month
  • API: ~€10/month
  • Total: ~€28/month

Quick Reference

  • 4 architecture steps
  • 5 epistemic labeling levels
  • 7 MCP Judilibre tools
  • ~€28 per month
View on GitHub Discuss this