Back

Contract Review

Auto-clear the known 80%, flag the unseen 20%.

Overview

Contract Review is a working AI workflow prototype for standard commercial contracts such as mutual NDAs and small order forms. The product idea is simple: if a clause matches a human-approved position, the system should clear it or redline it automatically. If the clause raises a genuinely novel issue, it should be escalated to a human with the reasoning already prepared.

The interesting part is not the chat interface. It is the operating model: approved positions become reusable memory, routine clauses stop consuming lawyer time, and new human decisions make the system better the next time a similar clause appears.

Scope Standard NDAs and order forms, not every possible contract.
Decision rule Auto-handle only when grounded in a human-approved position.
Stack Next.js, Vercel AI SDK, DeepSeek-compatible model routing, YAML position library.

Live demo

Live at clearnorm.com — preview embedded below.

Live preview clearnorm.com
Open clearnorm.com

Decision model

The system is built around issue-based positions, not contract-type templates. A limitation of liability clause, governing law clause, or confidentiality definition can be reviewed against the same approved position wherever it appears.

contract
   -> extract semantic clauses
   -> match each issue to an approved position
   -> grounded? auto_ok or auto_redline
   -> ungrounded? analyst opinion + adversarial refuter
   -> report: disposition, reasoning, redline, severity

Confidence is treated as grounding, not as a model score. If there is no approved position, the system does not invent one and act on it. It escalates the issue, drafts an analysis, and lets the human settle the position.

Evidence

Area Current evidence Honest claim
Clean standard NDA Project eval reports 11/11 clauses auto-handled with 0 false redlines. On standard paper, the lawyer should not read routine language.
NDA with deviations Eval catches 8/8 embedded deviations and escalates the novel non-solicitation issue. Known deviations can be pre-redlined; genuinely new issues still go to a human.
Architecture Small orchestrator in `src/lib/review.ts`; position library stored as reviewable YAML. The product is intentionally narrow, inspectable, and easy to extend by adding positions.
Workflow surface Web app supports sample loading, review jobs, position library, learning, settings, and assistant views. Portfolio-ready prototype, not a production legal SaaS.

Positioning

This is not a generic legal chatbot. It is a workflow product for a specific operational bottleneck: low-stakes standard contracts where most clauses are already familiar, but humans still spend time re-reading them.

Why it matters for AI product work: the system separates deterministic routing from model judgement, ties auto-decisions to approved human positions, and uses adversarial review only where the position library has no settled answer.