← Drew Bruce

Research & Writing

Working Papers

A small collection of working papers on AI architecture, governance, and verifiable accountability — the bits of my day job that refuse to fit into a blog post. These are drafts, not final word; I put them here because I’d rather have them read and argued with than polished and forgotten.

Comments, corrections, and strongly-worded disagreements are all welcome — andrewbruce@me.com.

Flagship Paper April 2026

The Observer Gap

Why AGI Won’t Emerge From Better Pattern Matching

Current approaches to artificial general intelligence rest on an implicit assumption: that human-level intelligence will emerge from sufficiently sophisticated information processing. This paper argues the assumption is backwards.

The missing component is what I call the Observer Function — a separable capacity to monitor cognition, represent it as an object of evaluation, assess it against higher-order goals, and modulate subsequent processing. Without it, you get better pattern matchers. You do not get general intelligence.

Read the PDF →
Companion Paper April 2026

The Observer Ecosystem

Governance as a Dynamic System of Competing Signals

A companion paper to The Observer Gap. The original paper treated the Observer Function as a single policy selector. Human experience suggests otherwise — curiosity pulls one way, caution another, empathy a third, and what we call a decision is the resolution of those competing signals.

This paper extends the model from a single observer to an ecosystem of balancing feedback loops, and sketches what that means for AI governance architecture.

Read the PDF →
Essay March 2026

Why human readiness is the missing layer in every AI strategy

On judgement atrophy, the identity shift, and what a real human-readiness layer looks like

Most enterprise AI strategies are built on two layers: technology and governance. Both get serious investment, serious executive air cover, and serious budget. There is a third layer that almost nobody is building deliberately — and it is the one that decides whether any of it delivers real value.

An argument for treating human readiness as a strategic layer in its own right, the two challenges nobody is talking about honestly, and what to invest in beyond training completion rates. Originally published on LinkedIn.

Read the essay →
Working Paper April 2026

The Trust Stack

A Protocol for Verifiable AI Accountability

As AI enters consequential decision-making — credit, medical triage, hiring, legal outcomes — existing compliance frameworks are inadequate. They rely on opinion, opaque models, and retrospective audit. The rating-agency failure is already being replayed.

This paper proposes a five-layer protocol combining cryptographic proof of process, distributed validation consensus, open-source audit logic, and outcome-agnostic incentives. Drawing on precedents from ICAO, SWIFT, Zero-Knowledge Proofs, and blockchain Proof of Stake — a path from “do I trust this company?” to “do I trust this protocol?”

Read the paper →
Companion Paper April 2026

The Assured Trust Stack

Stretching the Trust Stack across three classes of AI agent, with Internal Audit as the trust authority

The original Trust Stack assumed you owned every agent. Most enterprises don’t. By early 2026 the AI actually in use inside a mid-to-large organisation runs in three very different places — the agents you build, the agents embedded inside your SaaS vendors, and the agents your people invoke directly from their browsers. A trust layer that only covers the first covers a shrinking share of the problem.

This companion paper describes what the Trust Stack looks like when it has to stretch across all three, and names the body that sits at the top: Internal Audit in the primary design, a Big Four external attestor in the alternative. Includes detailed sequence diagrams for each class and a 90-day rollout plan.

Read the paper →
A note on status. These are working papers, not peer-reviewed publications. They represent my current thinking and are subject to revision. If you cite them, please note the draft date and check back for newer versions.