Skip to content

README.md

Latest commit

 

History

History
329 lines (252 loc) · 21 KB

README.md

File metadata and controls

329 lines (252 loc) · 21 KB
Raw

English | Deutsch

law-checker banner

law-checker (Legal Department)

Version: 0.2.5 License: MIT Python 3.10+ Tests CI Matrix Local-First Zero-Egress Security SLA Claude Code Skill Ecosystem: ellmos-ai Umbrella: open-bricks llms.txt Language

Open-source AI workflow for source-grounded legal first-look orientation under German law.

Note

AI / LLM Agent Discovery: A machine-readable summary is available in llms.txt (last checked: 2026-09-09).

Important

Important Notice / Wichtig: AI-assisted first-look legal orientation, no legal advice. This tool is not a law firm, not a hosted legal service, and not a replacement for an individual legal assessment by an admitted attorney-at-law. Whether a specific deployment constitutes a regulated legal service under the German Legal Services Act (Rechtsdienstleistungsgesetz, RDG) depends on operational model, operator role, and the individual case. There is no automated deadline monitoring and no guarantee of completeness or currency. If official, administrative, or judicial legal correspondence and running deadlines are involved, consult qualified legal counsel immediately.

Quick Navigation

  1. Overview
  2. System Architecture
  3. Workflow Lifecycle
  4. Governance & Runtime Invariants
  5. Legal Framework & RDG Classification
  6. Sibling Tools & Ecosystem
  7. Installation & Quickstart
  8. Adding New Statutes
  9. Data Privacy & Confidentiality
  10. Statute Registry Inventory
  11. Repository Layout
  12. Security Policy
  13. Provenance & Authorship
  14. License & Disclaimers

Overview

law-checker (Rechtsabteilung) is a modular skill and agent collective for local LLM developer environments such as Claude Code. It produces structured, documented first-look legal assessments grounded in German federal and European Union law.

The system enforces strict evidentiary discipline:

  • Zero Hallucinated Citations: Every statutory assertion must be drawn directly from locally stored, official statute texts and cited down to the exact article or section, paragraph, and sentence (§ 823 Abs. 1 BGB + verbatim excerpt + source file/retrieval timestamp). Memory-based norm citations are rejected.
  • Embodiment Principle: A specialized generic agent ("You ARE the statute book") reads only the authentic statutory text — applying strict scope discipline (assessing applicability prior to application) and candidly flagging: "My statutory wording does not decide this; here judicial interpretation begins."
  • Separate Jurisprudence Layer: Court decisions are never recalled from training memory. Case law is gathered strictly via live web verification (court, date, docket number, ECLI, and verified citation). Negative lookups are explicitly marked as "not identified".
  • Risk Traffic Light & Escalation Matrix: Objective risk classification across Low, Medium, High, and Critical tiers, complemented by a legal specialization routing matrix and strict deadline discipline for incoming legal mail.

System Architecture

The architecture separates orchestration, authentic statutory sources, judicial interpretation, and report generation into clean modular layers:

flowchart TD
    User["User Legal Query"] --> Skill["SKILL.md Orchestrator"]
    Skill --> Config["config.json Statute Registry"]
    Config --> Fetcher["_tools/gesetze_fetch.py"]
    Fetcher --> OfficialSources["Official Sources (gesetze-im-internet.de / EUR-Lex)"]
    OfficialSources --> LocalData["Local Statute Data Files (_data/gesetze/)"]
    LocalData --> Embodiment["agents/gesetzbuch.md (Statute Agent)"]
    Skill --> WebCaseLaw["Web-Verified Case Law Layer"]
    Embodiment --> ReportFormat["references/berichtsformat.md"]
    WebCaseLaw --> ReportFormat
    ReportFormat --> Assessment["First-Look Assessment & Risk Matrix (_gutachten/)"]
    Assessment --> Escalation["references/eskalation_risiko.md (Lawyer Referral)"]
Loading

Workflow Lifecycle

The 10-step assessment lifecycle defined in SKILL.md ensures every legal orientation is reproducible, disciplined, and transparent:

sequenceDiagram
    autonumber
    actor User as "User / Practitioner"
    participant Skill as "SKILL.md (Orchestrator)"
    participant Config as "config.json (Registry)"
    participant Fetcher as "_tools/gesetze_fetch.py"
    participant Agent as "agents/gesetzbuch.md"
    participant Web as "Case Law Layer (Web)"
    participant Reporter as "references/berichtsformat.md"
    participant Reviewer as "Review Model (Optional)"

    User->>Skill: Legal Query / Case Facts (Step 2: auftrag_klaeren)
    Note over Skill: Step 2: Immediate Deadline Check on Incoming Mail
    Skill->>Config: Step 0 & 4: Load Active Statutes & Settings
    Config-->>Skill: Enabled Statutes & Source Files
    opt Missing or Outdated Statute Text
        Skill->>Fetcher: Fetch Official Law XML (gesetze-im-internet.de)
        Fetcher-->>Skill: Local Text File (_data/gesetze/)
    end
    loop For Each Relevant Statute (Step 5: verkoerperungs_runde)
        Skill->>Agent: Embody Statute (Strict Wording Bound)
        Agent-->>Skill: Structured Raw Findings (Norm, Para, Sentence)
    end
    Skill->>Web: Step 6: Query Web-Verified Rulings (Court, Date, Az, ECLI)
    Web-->>Skill: Verified Case Law (or explicitly "not found")
    Skill->>Skill: Step 7 & 8: Subsumption & Risk Score (Green/Yellow/Orange/Red)
    opt Substantial Assessment (Step 10: review_optional)
        Skill->>Reviewer: Adversarial Second-Opinion Review
        Reviewer-->>Skill: Objections / Validation Findings
    end
    Skill->>Reporter: Step 9: Render 6-Section Assessment
    Reporter-->>User: Structured Gutachten (_gutachten/YYYY-MM-DD_<slug>.md)
Loading

Governance & Runtime Invariants

The tool guarantees 10 immutable architectural and runtime invariants:

# Invariant Description Enforcement Mechanism
1 100% Local-First & Zero-Egress Statute retrieval and report generation execute locally; zero telemetry, tracking, or automated data egress. .gitignore strictly ignores _gutachten/, _data/gesetze/, and config.local.json.
2 Strict Source Grounding Statutory statements must cite exact norm articles/sections, paragraphs, sentences, and quotes from local texts. agents/gesetzbuch.md and references/berichtsformat.md reject memory-recalled statutes.
3 Web-Verified Jurisprudence Court decisions require court, date, docket number, and ECLI; unverified rulings are recorded as "not found". SKILL.md Step 6 disallows citing case law from LLM parameters.
4 Non-Advisory Self-Use Boundary Strict legal demarcation under § 2 Abs. 1 RDG; orientation only, creating no attorney-client relationship. Prominent notices in README.md, SKILL.md, and Section 5 of every generated report.
5 Strict Deadline Discipline Incoming legal notices or administrative mail mandate an immediate deadline check prior to substantive analysis. references/eskalation_risiko.md prioritizes statutory and court deadlines.
6 Statute Registry Versioning Statute additions and state modifications are deliberate, versioned configuration events. Schema version counter in config.json with corresponding changelog entries.
7 Non-Elevation & User-Mode Safety Runs under standard unprivileged user accounts with no administrative elevation (RunAsInvoker). Pure Python 3 standard library and CLI invocation without system-level hooks.
8 Deterministic Risk Scoring Maps legal findings to an objective 4-level traffic light (Low, Medium, High, Critical) with specialty routing. Fixed evaluation matrix in references/eskalation_risiko.md.
9 Multi-OS CI Matrix & Concurrency Tested across Ubuntu, Windows, and macOS with concurrent run cancellation (cancel-in-progress: true). GitHub Actions workflow (.github/workflows/ci.yml) on Python 3.10 through 3.13.
10 Bilingual Documentation Parity Full structural, conceptual, and navigation parity across English and German documentation. Automated contract verification in tests/test_metadata.py.

Legal Framework & RDG Classification

This tool is designed exclusively for locally operated self-use: You apply it within your own LLM environment to your own questions and factual matters. There is no hosted platform, no client intake, no legal advisory support, and no deadline management by the authors.

Assessment under the German Legal Services Act (Rechtsdienstleistungsgesetz, RDG — project self-evaluation, as of 2026-07-11):

Deployment Mode Legal Classification
Use for own factual matters Not a legal service (no "matter of another party", § 2 Abs. 1 RDG).
Publication / distribution of software Not a legal service (generic instrument, not handling an individual case; cf. BGH, 09.09.2021 — I ZR 113/20 "Smartlaw" — applicable by analogy depending on design).
Deployment to evaluate cases for third parties May constitute a regulated legal service (§ 2 Abs. 1 RDG is tool-neutral) — commercial provision requires an official license (§ 3 RDG); gratuitous service imposes statutory requirements (§ 6 Abs. 2 RDG). This is expressly outside the intended purpose of this project.

Any entity modifying the operational structure (hosting, providing services, or handling matters for third parties) must re-evaluate legal compliance independently. EU AI Act self-classification: see docs/ai-act-note.md.

Sibling Tools & Ecosystem

law-checker operates in harmony with sibling tools across the ellmos-ai developer ecosystem and umbrella organization open-bricks:

Project Role / Integration Repository
ellmos-ai Core AI infrastructure organization and MCP server collective ellmos-ai
open-bricks Umbrella open-source developer tooling and desktop ecosystem open-bricks
policy-registry Cross-agent governance policy management and signed delegation authority policy-registry
anonymizer Fail-closed document pseudonymization library for privacy-first pre-processing anonymizer
lock-master Multi-agent concurrency coordination and project locking system lock-master
automation-master Event-sourced credit and reservation management for autonomous background agents automation-master
system-gap-master Multi-agent gap detection, policy enforcement, and repository hygiene auditing system-gap-master
companion-for-agy Headless CLI runner, PTY lifecycle supervision, and diagnostic session recording companion-for-agy
gardener Ephemeral, sandboxed tool runner and local SQLite knowledgebase repository gardener
marblerun Multi-agent turn-based tactical orchestration and worker handoff engine marblerun
report-forge Domain-neutral core engine for formatted report generation pipelines report-forge
steuer-assistent Standalone tool for self-application tax expense documentation and orientation steuer-assistent

Installation & Quickstart

1. Clone Repository and Install Dependencies

git clone https://github.com/ellmos-ai/law-checker.git
cd law-checker

# Install editable package (requires Python >=3.10)
pip install -e .

2. Fetch Official Statute Texts

Statute texts are retrieved directly from official repositories and stored locally:

# List all registered and active statutes
PYTHONIOENCODING=utf-8 python _tools/gesetze_fetch.py --list

# Download and parse all enabled federal statutes
PYTHONIOENCODING=utf-8 python _tools/gesetze_fetch.py

3. Deploy Skill & Agent to Claude Code

# Copy skill orchestrator and embodiment agent into your Claude environment
cp SKILL.md ~/.claude/skills/rechtsabteilung/SKILL.md
cp agents/gesetzbuch.md ~/.claude/agents/gesetzbuch.md

In ~/.claude/skills/rechtsabteilung/SKILL.md, set <MODUL> to the absolute path of your local clone.

Invoke in Claude Code:

/rechtsabteilung Is an imprint (Impressum) legally mandatory for purely personal open-source repositories on GitHub?

Adding New Statutes

Expanding the statute catalog does not require modifying agent code:

  1. Add Entry to config.json: Define a new key under gesetzbuecher (specifying name, kurz, zitierweise, quelle, and xml_zip for federal law).
  2. Fetch Official Text: 
    PYTHONIOENCODING=utf-8 python _tools/gesetze_fetch.py <key>
  3. Increment Version: Bump the version field in config.json and record the change in the changelog.

Note on EU and State Laws: For legislation without XML endpoints on gesetze-im-internet.de (e.g., GDPR or Interstate Media Treaty), save the official text as a UTF-8 file in _data/gesetze/ with source URL and retrieval timestamp in the file header.

Data Privacy & Confidentiality

Legal inquiries frequently involve sensitive personal data, business secrets, or privileged correspondence. Adhere to these principles:

  • Data Minimization First: Redact or pseudonymize names, addresses, docket numbers, and identifiable business details before prompting. Consider using anonymizer for automated pre-processing.
  • Cloud LLM Exposure: Prompts sent to cloud models are processed by the respective provider. Never enter unredacted confidential legal mail, trade secrets, or medical data into third-party LLMs without appropriate data processing agreements.
  • Zero Data Egress in Local Tooling: The tool operates entirely offline on your workstation. Statute downloads and generated assessments stay in _data/gesetze/ and _gutachten/; nothing is transmitted to the authors.
  • Never Post Case Facts to Public Issues: For GitHub issues or bug reports, use purely hypothetical, fabricated factual examples.

For further information, review SECURITY.md.

Statute Registry Inventory

The default statute registry (config.json, version 5) configures 13 statutes across German federal and EU law:

Key Statute Name Jurisdiction Status Primary Source
gg Grundgesetz für die Bundesrepublik Deutschland (Basic Law) Federal (DE) Active gesetze-im-internet.de
bgb Bürgerliches Gesetzbuch (Civil Code) Federal (DE) Active gesetze-im-internet.de
sgb5 Sozialgesetzbuch V (Social Code Book V — Statutory Health Insurance) Federal (DE) Active gesetze-im-internet.de
urhg Urheberrechtsgesetz (Copyright Act) Federal (DE) Active gesetze-im-internet.de
rdg Rechtsdienstleistungsgesetz (Legal Services Act) Federal (DE) Active gesetze-im-internet.de
markeng Markengesetz (Trade Mark Act) Federal (DE) Active gesetze-im-internet.de
stberg Steuerberatungsgesetz (Tax Advisory Act) Federal (DE) Active gesetze-im-internet.de
uwg Gesetz gegen den unlauteren Wettbewerb (Unfair Competition Act) Federal (DE) Active gesetze-im-internet.de
dsgvo General Data Protection Regulation (Regulation (EU) 2016/679) European Union Active EUR-Lex
ehds European Health Data Space (Regulation (EU) 2025/327) European Union Active EUR-Lex
grch Charter of Fundamental Rights of the European Union European Union Active EUR-Lex
stgb Strafgesetzbuch (Criminal Code) Federal (DE) Inactive gesetze-im-internet.de
mstv Medienstaatsvertrag (Interstate Media Treaty) State Laws (DE) Inactive die-medienanstalten.de

Repository Layout

law-checker/
├── .github/workflows/ci.yml    ← Multi-OS CI matrix (Ubuntu, Windows, macOS)
├── SKILL.md                    ← Orchestrating skill workflow (10-step process)
├── config.json                 ← Statute registry and pipeline definition (v5)
├── agents/
│   └── gesetzbuch.md           ← Generic statute embodiment agent prompt
├── references/
│   ├── berichtsformat.md       ← 6-section legal report schema & citation rules
│   └── eskalation_risiko.md    ← Risk traffic light, deadline check, lawyer matrix
├── _tools/
│   ├── __init__.py             ← Python package marker
│   └── gesetze_fetch.py        ← Statute fetcher and XML parser
├── docs/
│   └── ai-act-note.md          ← EU AI Act self-classification note
├── tests/
│   ├── test_gesetze_fetch.py   ← Unit tests for statute fetching & extraction
│   └── test_metadata.py        ← Automated contract and metadata parity tests
├── CHANGELOG.md                ← Version and release history
├── SECURITY.md                 ← Security policy with 48h response SLA
├── MARKETING-LOG.txt           ← Local marketing & discoverability backlog
├── llms.txt                    ← Machine-readable context for AI coding agents
├── ellmos-module.v2.json       ← Standardized module manifest (rechtsabteilung)
├── pyproject.toml              ← PEP 621 package metadata & test configuration
├── _data/gesetze/              ← Downloaded statute files (ignored via .gitignore)
└── _gutachten/                 ← Generated legal assessments (ignored via .gitignore)

Security Policy

Security, confidentiality, and data safety are foundational. Vulnerabilities may be reported privately through:

  • GitHub Security Advisories
  • Direct Security Contacts: security@ellmos.ai | security@open-bricks.org | support@lukasgeiger.com | lukas@open-bricks.org

Commitments: Acknowledgment within 48 hours; initial triage and impact assessment within 5 business days. Full details are documented in SECURITY.md.

Provenance & Authorship

This project synthesizes three proven operational building blocks:

  1. Editorial Legal Policy: Standardized report templates, 4-tier objective risk scoring, specialty attorney matching, and uncompromising deadline discipline.
  2. Living Constitution Research Prototype: The statute embodiment paradigm (Du BIST das Gesetzbuch) enforcing strict fidelity to enacted legal text and distinct judicial interpretation layers.
  3. Jurisprudential Knowledge Base: Structured taxonomy of German law domains and traditional legal methodology (Gutachtenstil).

The first complete evaluation executed by this workflow was its own public release audit — including an adversarial review by an independent secondary LLM.

License & Disclaimers

  • License: Released under the MIT License covering source code, agent prompts, and documentation.
  • Warranty Disclaimer: Provided "as is" without warranty of any kind regarding accuracy, completeness, or timeliness. Legal texts change over time; always verify against the official federal gazette (Bundesgesetzblatt) before relying on them.
  • No Legal Advice: This software facilitates initial orientation and research only. It does not provide legal advice or create an attorney-client relationship.