research
Papers & Publications
Jun 23, 2026 · 46 min read
paperThe Sovereign Memory Layer
The capstone of The Human Layer series. The durable layer of an AI-native organization is the memory it owns and the judgment it can prove, not the model it rents. This paper names continuity as the objective and memory as the mechanism, introduces decision lineage as its atomic unit, and shows why memory has crossed from feature to infrastructure, the layer that survives when everything above it is rented, swapped, or withdrawn.
May 9, 2026 · 43 min read
paperThe Human Layer Economics
The fourth and final paper in The Human Layer series. If the evidence for augmentation over automation is this consistent, why does capital continue to flow toward replacement? This paper answers that question through the lens of incentive architecture: venture capital pricing, AI vendor pricing models, and executive performance structures that systematically make human layer elimination appear rational while externalizing the costs. It introduces organizational automation bias as a formally delimited concept, documents the three forces now collapsing the temporal liability gap that enables it, and makes the empirical case that the Human Layer Score is positioned to function as a capital markets signal for organizations operating in regulated environments.
Apr 7, 2026 · 41 min read
paperThe Human Layer Audit: Measuring Accountability in AI Systems
An architecture that cannot be measured cannot be enforced. This paper operationalizes the Human Layer into a scoring and assessment framework: a risk-tiered maturity model, identity anchoring requirements, compliance traceability, and a floor-based scoring methodology designed to prevent organizations from averaging away weaknesses. Builders use it to design. Auditors use it to evaluate. Regulators use it to define what human oversight must actually mean in practice.
Mar 19, 2026 · 23 min read
paperThe Human Layer Architecture: A Specification for Human-AI System Design
Paper 1 established why the Human Layer matters. This paper defines how to build it. Five architectural components, each independently required by emerging regulation, none previously unified into a single specification. Decision gates, escalation protocols, accountability structures, override mechanisms, and trust calibration interfaces form a dependency graph where removing any component breaks the guarantees provided by the others.
Feb 24, 2026 · 21 min read
paperThe Human Layer: Why the Most Critical Infrastructure in AI Isn’t Artificial
As organizations race toward full automation, many are removing the very layer that determines whether AI systems succeed at scale: the human one. This paper introduces the Human Layer as architectural infrastructure, not philosophical preference. Drawing on economic data, historical precedent, and real-world system design in regulated environments, it argues that AI systems built to amplify human judgment consistently outperform those designed for pure replacement. In trust-dependent, regulated, and ambiguity-rich domains, human oversight is not overhead. It is load-bearing. This publication establishes the case for why the Human Layer must be designed as a structural component of any serious AI system and introduces an audit framework to evaluate whether it truly exists.
The Human Layer Framework
Five components
Each is an architectural requirement rather than a practice, and each is scored 0 to 4 across three consequence tiers.
- Decision gates
- A point at which the system must stop and obtain a human decision before proceeding.
- Escalation protocols
- Defined routes by which a system hands a decision upward when it exceeds its authority.
- Accountability structures
- Mappings that connect system outcomes to the responsible human.
- Override mechanisms
- Means by which a human retains authority over consequential actions the system proposes.
- Trust calibration interfaces
- Surfaces where a system's outputs are validated against real-world context.
Questions
How do you measure whether human oversight of an AI system is real?
You measure it the way you would measure any other piece of infrastructure: by naming the components that must exist and scoring each one. The Human Layer Framework specifies five, and scores each from 0 to 4 across three consequence tiers. An oversight arrangement that cannot produce a score is an intention, not a control.
What is the difference between a human in the loop and meaningful human oversight?
A human in the loop describes a position in a workflow. Meaningful oversight describes an architecture: a decision gate that halts the system, an escalation route when authority is exceeded, an accountability mapping to a named person, an override that works, and an interface that lets the human calibrate trust. A human can be in the loop while none of those exist.
How do you tell rubber-stamping from genuine review?
Rubber-stamping is what happens when a decision gate exists but the surrounding components do not. If the reviewer has no override that carries authority, no escalation route, and no interface showing why the system reached its output, approval is the only available action. The framework treats that as a scoring failure rather than a human failure.
What should an AI oversight maturity assessment contain?
It should score each oversight component separately rather than producing one overall rating, and it should vary the standard by consequence, since the same system may warrant light oversight in one use and strict oversight in another. The Human Layer Audit scores five components across three consequence tiers for that reason.
How does the Human Layer Framework relate to the EU AI Act?
Article 14 of the EU AI Act requires that high-risk AI systems be subject to effective human oversight. The Act states the obligation but does not specify an implementation. The Human Layer Framework is one way to make oversight concrete and measurable: five named components, each scored, so an organisation can show what its oversight consists of rather than assert that it exists.