LIA: A Scope-Contract Control Plane for Surgical, Auditable Contract Editing - A Systems Architecture and Prospective Evaluation Protocol

Abstract

AI contract-editing systems often replace whole clauses where expert redlines use smaller, distributed tracked changes. We call this descriptive difference the surgicality gap. It may arise from model judgment, the editing harness, or both. This paper presents LIA, a proposed controlplane architecture, and a prospective protocol for estimating the harness contribution while preserving legal sufficiency. LIA adapts computer-assisted-translation concepts-segmentation, precedent memory, match-rate leverage, and banded routing-and applies the narrower security principle that a model may propose an action while code outside the model determines whether the action is permitted. The architecture uses a hierarchical document map with versioned structural anchors and cross-revision reconciliation; a leverage pass that separates reuse similarity from policy-based action eligibility; hard-veto features for predefined high-risk differences; scope contracts enforced at the only document-write boundary; run-scoped provisional memory; and a fresh-session whole-document audit that cannot retrieve the current run's unapproved outputs. Its intended guarantee is limited and mechanical: edits that violate an expressible scope rule are rejected before document mutation. It does not guarantee legal correctness, complete detection of meaning reversal, or immunity from prompt injection. We contribute the architecture, a scope-contract schema and enforcement model, provenance-aware diagnostics, and a three-arm evaluation design with falsifiable morphology and sufficiency hypotheses. LIA is under active implementation, and this paper reports no experimental results.