# Legal AI Reliability

We publish a public, lawyer-first taxonomy of AI failure modes in contract drafting and editing. This is a reliability library that makes failure modes concrete, shows why they matter, and documents how they are detected.

## Why This Exists

Lawyers and clients need reliability before automation. Our experience is that models still violate document invariants that trained lawyers routinely protect: numbering integrity, defined-term consistency, cross-reference accuracy, and minimal-diff editing discipline.

## Who This Helps

### In-house Teams

A grounded way to explain why AI still needs human review in contracts, without sounding anti-innovation.

### Foundation Labs

A structured, reproducible failure-mode library that can be turned into evals, lint rules, and training data.

## Failure Mode Taxonomy (Draft)

*   Document structure invariants (list nesting, parent-child slicing)
*   Defined-term integrity (capitalization, duplication, drift)
*   Cross-reference integrity (broken sections, exhibit callouts)
*   Redline mechanics (non-minimal diffs, numbering cascades)
*   Obligation logic drift (shall/may flips, condition changes)
*   Citation and quote verification (false citations, wrong paragraph)

## Taxonomy Index

Each item links to the gallery entry with a short, lawyer-first explanation.

| Category | Failure Mode | Severity | Detectable | Status |
| --- | --- | --- | --- | --- |
| Structure | List Integrity (Orphaning) | High | Yes | Lintable |
| Structure | Renumbering Cascade | High | Yes | Lintable |
| Definitions | Definition Alphabetization | High | Yes | Lintable |
| Definitions | Defined Term Capitalization | High | Yes | Lintable |
| Cross-References | Dirty Cross-References | High | Yes | Lintable |
| Comparisons | Distractor Misattribution | Critical | Architectural | Architectural |
| Citations | Unread / Stale Citations | High | Yes | Lintable |
| Identifiers | Identifier Interpolation | High | Yes | Lintable |
| Parties | Placeholder & Party Consistency | High | Yes | Lintable |
| Semantics | Semantic Overcompression | Medium | Partial | Partial |
| Formatting | Formatting & Context Overhead | Medium | Architectural | Architectural |
| Markup | Visual Reasoning / Redlines | High | Architectural | Architectural |
| Citations | TOC Over-Indexing | Medium | Yes | Lintable |
| Editing | Document vs Code Editing | High | Architectural | Architectural |

## What Comes Next

We are building public examples and a private evaluation pack. If you want the dataset, scoring harness, or a lab-facing briefing, contact us.

[Contact UseJunior](/contact.md)