The body of work.

Architecture-Sensitive Duties: The Convergence of Emerging Bar Doctrine and Bounded Legal-AI Architecture

Best for readers interested in legal ethics, doctrine, and the architectural conditions of dischargeable duties.

The legal profession’s duties — competence, candor, confidentiality, and supervision — do not change because lawyers begin using artificial intelligence. The conditions under which those duties can be realistically discharged do change. The paper introduces the doctrinal concept of architecture-sensitive duties: duties whose practical dischargeability depends on the architectural properties of the AI system in use, in ways that cannot be remedied by care alone. It identifies the architectural conditions on which independent judgment and confidentiality most directly depend, and shows how emerging bar doctrine is converging on the same structural concerns from a regulatory direction.

Evidence Intelligence and the Contamination Problem in Legal AI: Architectural Preconditions for Trustworthy Legal AI

Best for readers interested in the evidentiary substrate, contamination, and whole-record reasoning.

Two structural contaminations afflict legal-AI systems organized around chat or fragmented retrieval: first-contact contamination, in which user framing reaches the analytical layer before independent reading has occurred; and contamination at the join, in which the record exceeds what the system can hold and is reconnected by prompts. The paper identifies the conjunctive architectural response — touchless ingestion paired with whole-record reasoning — and the supporting conditions on which trustworthy first-pass legal analysis depends.

The Five Hallucination Mechanisms

The legal profession is using one word — hallucination — to name several different architectural failures. The essay names five distinct mechanisms behind that one word: first-contact contamination, session-context contamination, join-by-prompt contamination, source-claim detachment, and training-prior fabrication. Three are structurally excluded by a bounded analytical architecture. One is constrained. One is reduced. The right cross-architecture question is not how much a system hallucinates in the aggregate, but which mechanisms it admits at all.

Foundations Don't Have Plumbing

A foundation in construction is reinforced, continuously poured, and not penetrated by later trades. A slab is intentionally jointed, accessible, and built to accept modification. The essay carries that distinction into legal AI: workflow products are slabs and are supposed to be; the analytical layer must be a foundation, and a foundation that admits integrations after the pour is no longer a foundation.

A Handoff Is a Surface

A handoff from evidence analysis into drafting is already a surface in the relevant architectural sense, even when it never leaves the vendor’s stack. The relevant boundary is not corporate. It is functional. The question is not who owns the next tool, but whether the analytical layer can reach it at all.

Methodology Is Not Prompt Stain

Not every frame is a stain. This paper answers the objection that every analytical system has a frame somewhere by distinguishing methodology from hypothesis and proposing a four-property test for determining whether a frame is a non-contaminating analytical discipline or a staining input.

Prompt Stain: Why Provenance Cannot Survive the First Prompt

This paper proposes the term prompt stain for the loss of first-read provenance that occurs when user framing reaches the analytical layer before the system has independently read the evidence. It argues that the problem is structural rather than procedural and explains why better prompting cannot restore an unstained read.

The First Contamination

When a user describes a matter to an AI system before the system has read the evidence, the analytical operation has already been shaped by that framing. The loss of first-read independence is structural, not procedural — no amount of careful prompting can recover what would have been seen.

The Second Contamination

When a record exceeds what a system can hold, something has to do the joining. If the join is supplied by user prompts, retained instructions, or retrieval rules, the system is no longer reasoning from the record alone. The second contamination names that loss and identifies whole-record reasoning as the architectural response.

The Architecture and the Attorney

Analysis and judgment are different kinds of work. The architecture produces evidence-tethered analysis at scale; the lawyer brings judgment about which theories to plead, what to leave out, and what to test first. Conflating them is the design error this essay names.

The Bounded System

A boundary on a diagram is not a perimeter in a system. The same architectural property that protects the lawyer’s judgment also protects the client’s confidences: integrity and security are one commitment, looked at from two angles. Reaching them from an open-surface system is not a code change. It is a category change.

The Lawyer’s Duties

Competence, candor, confidentiality, and supervision are stable. The conditions under which they are discharged are not. This essay walks the four duties most clearly affected by AI architecture and identifies, for each, the architectural condition on which its practical discharge increasingly depends.