Tribunal Harness

A UK employment tribunal legal intelligence engine for litigants-in-person (LiPs). Tribunal Harness addresses the information asymmetry between unrepresented claimants and respondents with legal representation, providing schema-driven legal analysis across the full range of employment tribunal claim types.

This is not a chatbot. The core architecture is built on four pillars: Inverse Chatbot (dynamic schema-driven UI instead of free chat), Epistemic Quarantine (every legal proposition carries a trust level; ungrounded claims are stripped), Durable State Machine (long-running case state), and Adversarial Shadow-Opponent (a Drafter/Critic/Judge debate that stress-tests arguments before the user sees them).

Important: This tool provides legal information, not legal advice. Every user-facing output carries a persistent disclaimer in accordance with the Legal Services Act 2007.

What's implemented

• 10 claim-type schemas (one file each) — unfair dismissal, direct/indirect discrimination, harassment, victimisation, reasonable adjustments, whistleblowing, wrongful dismissal, fire & rehire, zero-hours rights.
• ERA 2025 awareness — all commencement dates centralised in src/lib/constants.ts; a live tracker drives the UI.
• Deadline calculator — handles the 3-month/6-month time-limit regime change, UK bank holidays, and ACAS Early Conciliation extensions.
• Live case-law lookup — authorities are retrieved and verified live from TNA Find Case Law (free, no key), not a pre-built RAG corpus. /api/analyse double-checks every AI-cited authority and marks it VERIFIED only on an exact neutral-citation match, falling back to a curated list of pre-2003 landmarks. See tribunal-harness/docs/live-case-law.md.
• 3-agent debate engine (/api/debate) — Drafter → Critic → Judge, scored on a 100-point rubric.
• Legal-writing refinement — LLM responses are post-processed to polish prose while preserving the JSON schema.
• Compliance infrastructure — UK GDPR Article 9 explicit-consent gate, LSA 2007 disclaimers on every page.
• PDF → Markdown — source PDFs are converted to clean Markdown before any LLM reasons over them.

Tech Stack

The repository contains two layers. The tribunal-harness/ Next.js app is the current production build; the root-level Vite project is an earlier prototype kept for reference.

Current build (tribunal-harness/):

• Framework: Next.js 15.1.0 (App Router, Turbopack), TypeScript strict mode
• UI: React 19, Tailwind CSS 4, Framer Motion, lucide-react
• AI: Anthropic SDK (@anthropic-ai/sdk) — hub-and-spoke routing (Haiku → triage, Sonnet → analysis, Opus → critic/judge). All LLM routes degrade gracefully without an API key.
• Document parsing: mammoth (DOCX), pdf-parse (PDF)
• Testing: Vitest (unit/integration); a hermetic end-to-end smoke harness; ESLint

Earlier prototype (root src/ — Vite/React 18, JSX).

How to Run

cd tribunal-harness
npm install
npm run dev          # dev server (Turbopack) → http://localhost:3000

npm run build        # production build (must pass with 0 errors)
npm test             # Vitest — 215 tests
npm run lint         # ESLint
npm run smoke        # hermetic end-to-end run, no API key required (see below)

Offline / no-API-key mode

Set LLM_PROVIDER=agent to route every LLM call to a deterministic, schema-conformant stand-in (src/lib/llm/agent-provider.ts). npm run smoke uses this to invoke every API route in-process and write smoke-report.{json,md} — letting the full pipeline (including the debate engine) run with no Anthropic key. With a real key, set ANTHROPIC_API_KEY in tribunal-harness/.env.local (see .env.example).

Project Structure

tribunal-harness/                # repo root (Vite prototype lives here)
├── tribunal-harness/            # ← current build, Next.js App Router (primary)
│   ├── src/
│   │   ├── agents/              # LLM system prompts with rationale comments
│   │   ├── app/                 # pages + API routes (analyse, triage, debate,
│   │   │                        #   deadlines, case-law/{find,judgment,search}, …)
│   │   ├── components/          # React components
│   │   ├── lib/                 # constants.ts (ERA 2025 SSOT), claude-client/-config,
│   │   │                        #   verified-authorities, llm/ (agent stand-in)
│   │   ├── schemas/             # 10 claim-type schemas as TypeScript interfaces
│   │   ├── services/            # deadline-calculator, find-case-law, citation-validator,
│   │   │                        #   legal-writing-refinement, pdf-to-markdown
│   │   └── types/               # shared TypeScript types
│   ├── scripts/smoke-run.ts     # end-to-end smoke harness
│   ├── docs/                    # live-case-law.md, enterprise-audit.md
│   └── CLAUDE.md                # app-level context (read this first)
│
├── src/                         # earlier Vite/React prototype
├── corpus/authorities/          # case-law manifest (judgment PDFs are re-fetchable, gitignored)
├── Design/                      # UI design assets
└── CLAUDE.md                    # full project context and coding standards

Status

In active development. The schema-driven analysis engine, deadline calculator, ERA 2025 tracker, live case-law lookup, 3-agent debate engine, Noir design system, and compliance infrastructure (GDPR, LSA 2007) are implemented and covered by 215 passing tests. Open items: a managed API layer to replace BYOK as the default data flow (required before institutional pilots), and the durable Temporal.io state machine (Phase 4).