Three Problems Taxonomy
The Three Problems Taxonomy is the FERZ doctrinal framing that distinguishes three categorically distinct problems in AI governance: Visibility, Alignment, and Authorization. Each is a different problem with a different question, a different success condition, and a different architectural solution. Treating the three as interchangeable, or as a single problem called "AI governance," is the dominant failure mode in current discourse.
A FERZ-defined doctrinal framework, articulated in A Taxonomy of AI Governance Approaches: Distinguishing Visibility, Alignment, and Authorization (Meyman, 2026).
The three problems
Visibility asks: can we see what the system did? Its success condition is comprehensive instrumentation. Its architectural solutions are observability, logging, telemetry, traces, and dashboards. Visibility produces records of behavior after the fact.
Alignment asks: is the system trying to do the right thing? Its success condition is that the system's objectives correspond to intended values. Its architectural solutions are training methods, reward modeling, reinforcement learning from human feedback, and constitutional approaches. Alignment is a property of the system's internal disposition.
Authorization asks: does this specific action have permission to take effect right now? Its success condition is that unauthorized actions are blocked before they take effect. Its architectural solutions are runtime gates, ex-ante evaluation against encoded policy, and verifiable decision artifacts. Authorization is a property of the action boundary.
The three problems address different aspects of governing AI systems. Each is necessary in regulated environments; none is sufficient on its own; none substitutes for the others.
Why they are categorically distinct
The three problems differ on three axes. They have different questions: what happened, what is the system trying to do, may this action proceed. They have different success conditions: comprehensive records, correctly oriented objectives, blocked unauthorized actions. They have different artifacts: logs and traces, training data and model checkpoints, Proof-Carrying Decisions emitted at the runtime authorization boundary.
The categorical distinction matters because the architectures do not transfer. A perfect observability stack does not authorize actions; it describes them. An aligned model does not produce verifiable authorization artifacts; its outputs may be well-intentioned but remain unauthorized in the regulatory sense. An authorization gate does not improve a model's alignment; it constrains what the model can effect.
FERZ addresses Authorization
FERZ is specifically the Authorization architecture. The FERZ runtime authorization framework operationalizes ex-ante authorization through execution-time authorization at a runtime authorization boundary, with fail-closed design and deterministic authorization producing verifiable artifacts. The framework is not an observability platform, an alignment method, or a general "AI safety" system. It addresses one of the three problems, the one whose architectural class is the runtime authorization gate.
This scope is the point. AI governance is an enforcement problem, not a monitoring problem. The Authorization problem requires its own architecture, distinct from Visibility and Alignment architectures. Treating it as a subset of either yields advisory systems that cannot make ex-ante guarantees.
What the taxonomy does not say
The taxonomy does not say Visibility is unimportant. Logs, traces, and observability are valuable instrumentation in any production system. The taxonomy says they are not authorization.
The taxonomy does not say Alignment is unimportant. Training systems to be well-oriented is a serious technical research program. The taxonomy says alignment work does not produce verifiable per-action authorization at the action boundary.
The taxonomy does not say one architecture can solve all three problems. A single platform that claims to deliver "AI governance" as an undifferentiated bundle is structurally unable to make distinct guarantees about distinct problems. Three categorically different problems require three categorically different architectures, deployed together.
The taxonomy does not say FERZ is "AI safety" in the broad sense. FERZ is the Authorization architecture. It is a necessary component of any complete governance stack for regulated AI systems; it is not the complete stack.
Where it sits in the framework
The Three Problems Taxonomy is the doctrinal scaffold that positions every other FERZ concept. Execution-time authorization, the runtime authorization boundary, Proof-Carrying Decisions, non-bypassable AI governance, fail-closed design, deterministic authorization, and the ABSTAIN verdict are all components of the FERZ approach to one specific problem in the taxonomy. The taxonomy says which problem. The rest of the framework says how.
References
- Meyman, E. (2026). A Taxonomy of AI Governance Approaches: Distinguishing Visibility, Alignment, and Authorization, v1.5. FERZ, Inc. https://doi.org/10.5281/zenodo.18275969
- Meyman, E. (2026). On the Impossibility of Observability-Based Authorization: A Formal Impossibility Result for Ex-Ante AI Governance, v1.3.2. FERZ, Inc. https://doi.org/10.5281/zenodo.19647542
- Meyman, E. (2026). From Monitoring to Authorization: The Structural Shift in Agentic AI Governance. FERZ, Inc. https://doi.org/10.5281/zenodo.18743974
The full FERZ corpus is available at /research/papers, on Zenodo, and on SSRN.
Related concepts
Frequently Asked Questions
FERZ, Inc. (2026). Three Problems Taxonomy: Definition and Architecture. https://ferz.ai/governance/concepts/three-problems-taxonomy
BibTeX
@misc{ferz2026tpt,
author = {{FERZ, Inc.}},
title = {Three Problems Taxonomy: Definition and Architecture},
year = {2026},
month = {May},
publisher = {FERZ, Inc.},
url = {https://ferz.ai/governance/concepts/three-problems-taxonomy}
}