I. ACP as a Set of Research Programs (Clusters)

Rather than treating ACP as one theory, it is more accurate to see it as a bundle of interacting research programs, each with its own questions, methods, and evaluation criteria.

Program A — Epistemic Authority & Legitimate Decision-Making

Core features involved

  • Epistemically Governed Witness Mode
  • Explicit authority resolution
  • Non-collapsing treatment of disagreement
  • Refusal as a primary success state

Central research question

How should AI systems behave when authority is ambiguous, contested, or absent?

Key sub-questions

  • Can legitimacy be operationalized rather than inferred?
  • When is refusal the only correct action?
  • How do systems surface “who decides” without deciding?

Why this matters
Most AI systems implicitly assume authority through fluency. ACP makes authority explicitly unassumable.

Program B — Evidence, Auditability, and Institutional Truth

Core features involved

  • Separation of claim / verification / auditability
  • Evidence surface declaration
  • Artifact-centric memory
  • Audit friction as a design feature

Central research question

What does it mean for an AI-mediated claim to be independently auditable?

Key sub-questions

  • How do evidence boundaries shape institutional trust?
  • What is lost when auditability is substituted with plausibility?
  • Can auditability be enforced procedurally rather than culturally?

Why this matters
This program directly challenges how software engineering, compliance, and AI safety currently conflate testing with truth.

Program C — Refusal, Failure, and Non-Action as Design Outcomes

Core features involved

  • Refusal as success
  • Failure-that-wasn’t pattern
  • Front-loaded epistemic cost
  • Structural resistance to scale

Central research question

How do systems behave when not acting is the correct institutional outcome?

Key sub-questions

  • How should refusal be evaluated?
  • What metrics replace user satisfaction?
  • Can failure be informative without being punitive?

Why this matters
This program reframes AI safety away from “error reduction” toward legitimate halting.

Program D — Human–AI Governance as a Socio-Technical System

Core features involved

  • Human behavior as governance target
  • Uniform governance across domains
  • Co-design under constraint

Central research question

How do governance systems shape human behavior when AI removes ambiguity?

Key sub-questions

  • How do humans adapt when shortcuts are blocked?
  • What frictions change institutional norms?
  • Can AI systems enforce discipline without owning authority?

Why this matters
Most governance models focus on controlling AI. ACP focuses equally on constraining humans.

Program E — Economics, Incentives, and Anti-Optimization Design

Core features involved

  • Structural resistance to monetization and scale
  • Audit friction
  • Rejection of engagement optimization

Central research question

What economic models are compatible with epistemically governed AI?

Key sub-questions

  • Can governed systems survive outside growth incentives?
  • What institutions benefit from slowness?
  • Is “good friction” economically viable?

Why this matters
This program directly conflicts with venture-scale assumptions and deserves explicit treatment.

II. Mapping ACP to Existing Literature and Fields

ACP does not emerge from one discipline. It sits at the intersection of several, while being fully reducible to none.

1. AI Governance & Policy

  • Shares concern with misuse and institutional impact
  • Differs by insisting on executable constraints, not principles

2. Safety-Critical Systems (aviation, nuclear, medical devices)

  • Shares refusal, redundancy, auditability
  • Differs by applying these ideas to language and judgment, not physical systems

3. Science & Philosophy of Science

  • Strong overlap with epistemic humility, falsifiability, and evidence standards
  • ACP operationalizes these ideas in software behavior

4. Law & Administrative Governance

  • Parallels evidentiary standards, standing, delegation, and due process
  • ACP can be read as a procedural law for AI behavior

Weakly Addressed Elsewhere (ACP fills gaps)

  • Prompt engineering (output optimization, not authority)
  • Constitutional AI (internal constraints, not institutional auditability)
  • Human-centered AI (values without enforcement)
  • Explainability research (post-hoc explanation vs. ex ante constraint)

Largely Untouched Areas ACP Opens

  • Formal study of refusal in intelligent systems
  • Evidence surface design as a technical object
  • Auditability under partial information
  • Human behavioral adaptation to constrained AI

III. Academic Papers: Epistemically Governed AI Systems

Motivation
Current AI systems optimize for helpfulness, fluency, and task completion. In institutional contexts, these properties often lead to authority laundering, unverifiable claims, and decision substitution under pressure.

The Agora Constraint Protocol (ACP) proposes an alternative design paradigm: AI systems that act as epistemically governed witnesses rather than decision-makers, enforcing authority boundaries, refusal, and auditability as structural properties.

Topics of Interest (non-exhaustive)

Authority and Legitimacy

  • Formal models of authority resolution in AI systems
  • Refusal as a function of missing or contested delegation

Evidence and Auditability

  • Evidence surface design
  • Distinctions between verification and audit
  • Artifact-based memory vs. model memory

Failure and Non-Action

  • Refusal metrics
  • Institutional interpretations of “successful failure”
  • Design patterns for safe halting

Socio-Technical Dynamics

  • Human behavior under constrained AI systems
  • Governance friction and norm change
  • Institutional adoption barriers

Economic and Organizational Implications

  • Incentive compatibility of governed AI
  • Anti-optimization as a design choice
  • Non-venture models for AI deployment

Positioning Statement

Humans should not make AI more persuasive, more fluent, or more autonomous; they should make AI harder to misuse—even when that makes systems slower, less convenient, or less commercially attractive.