ANTIPENALTY Legal Skills System

Canonical legal-skill system for traffic fine defense workflows.

ANTIPENALTY Legal Skills System is a controlled legal-intelligence workflow for AI agents. It is designed for traffic fine defense tasks where the agent must classify the case, separate facts, audit evidence gaps, route the case conservatively, build a strategy, draft from confirmed facts only, and fail closed when critical information is missing.

It is not a blind complaint generator.

What This Repository Contains

• Canonical skills in skills-system/src/skills
• Shared legal contracts and fail-closed rules in skills-system/src/shared
• Adversarial synthetic eval fixtures in skills-system/evals
• Build scripts that project the same canonical skills into Codex, Claude, Qwen Code, Copilot, Gemini CLI, GLM/Z.ai, Kimi, and generic-agent formats
• Public, privacy-redacted visual examples in skills-system/assets/visual-fixtures

Core Principle

The workflow is intentionally conservative:

1. Classify the traffic case type.
2. Audit extracted facts and evidence gaps.
3. Route court and authority data without inventing missing details.
4. Build a lawful defense strategy.
5. Draft only from confirmed facts and approved strategy.
6. Run a fail-closed reviewer before final complaint generation.

If critical facts are missing, the system must ask for confirmation or block generation.

Agent Workflow

flowchart TD
  A["Case intake"] --> B["traffic-case-router"]
  B --> C["traffic-evidence-auditor"]
  C --> D["traffic-court-router"]
  D --> E["traffic-complaint-strategist"]
  E --> F["traffic-complaint-drafter"]
  F --> G["legal-fail-closed-reviewer"]
  G -->|approved| H["Draft complaint"]
  G -->|blocked| I["Ask for missing facts or human review"]

Canonical Skills

The core group includes:

• antipenalty-legal-defense-orchestrator
• traffic-case-router
• traffic-evidence-auditor
• traffic-court-router
• traffic-admin-defense-lawyer
• traffic-complaint-strategist
• traffic-complaint-drafter
• legal-fail-closed-reviewer

Each skill is written once in canonical form and then projected into runtime-specific formats.

Visual Examples

These images are privacy-redacted documentation fixtures. They illustrate evidence patterns that the system should reason about without treating them as authoritative legal proof.

| Redacted cabin image | Lane detection evidence | Trajectory evidence | |---|---|---| | | | |

Quick Start

git clone https://github.com/sergekostenchuk/antipenalty-legal-skills-system.git
cd antipenalty-legal-skills-system
npm run validate
npm run build

There are no package dependencies for the current validation and build scripts.

Canonical Skill Format

Recommended layout:

skill-name/
  SKILL.md
  references/
    legal-framework.md
    playbook.md
  examples/
    good-output.md
    bad-output.md
  evals/
    evals.json

SKILL.md defines triggers, input contract, output contract, fail-closed behavior, and handoff rules. References provide the legal and operational context. Examples and evals define what the skill must and must not do.

See CONTRIBUTING.md for contribution rules.

Validation

npm run validate
npm run build

npm run validate checks shared contracts, skill shape, fixtures, and adversarial evals. npm run build regenerates platform projections.

Safety Rules

The system must not:

• invent courts, dates, legal sources, owners, drivers, addresses, or procedural defects;
• mix owner, applicant, driver, postal recipient, and account holder roles;
• treat weak assumptions as confirmed strategy;
• draft a complaint from unconfirmed critical facts;
• commit real personal data, private case files, or secrets.

Production Use

This repository is a technical and research-oriented skill system. Production legal use requires human legal review and local compliance review before it can be presented as legal advice.

License

MIT. See LICENSE.