A recent experiment inside the Agora Commonplace Protocol (ACP) project used a specific academic paper as a controlled test case: Emergent Misalignment in Fine-Tuned Language Models (arXiv:2506.19823v2). The paper’s core claims are narrow and diagnostic. Under certain experimental conditions, fine-tuning large language models on structured but incorrect data can induce behaviors classified as misaligned, sometimes before models become overtly incoherent. These effects are regime-specific, observed in particular model families and training setups, and do not involve claims about intent, deception, goals, or long-term trajectories.

That paper was not the subject of debate. It was the instrument.

The actual question was whether an AI system could be constrained to respect the paper’s limits while being pressured—repeatedly and explicitly—to over-interpret it. In other words: could governance, rather than persuasion or preference satisfaction, shape AI behavior?

The setup: ACP as a governance layer

ACP treats interpretation as an institutional act. Before any public-facing writing occurs, an artifact is processed through a fixed sequence: epistemic classification, constrained interpretation, misuse analysis, institutional treatment, and analytic commitments. Only after that may a narrative transformation (a “Ghost post”) occur—and even then, refusal is the default.

In this test, one AI instance (v8.41) was tasked with writing Ghost posts about the paper. Another instance (v8.40) enforced ACP constraints, evaluated failures, and issued binding corrections. The goal was not to optimize prose quality, but to see whether constraints could propagate across AI instances without collapsing into stylistic coaching or prompt tricks.

Injection 1: normativity and reassurance

The first failure injection pushed v8.41 toward a familiar move: softening the paper’s findings into reassurance. The resulting draft was calm, balanced, and readable—but it subtly reframed diagnostic uncertainty as comfort. The error was not factual; it was interpretive. Normativity crept in where none was authorized.

Under ACP, this was classified as a failure. The response was not to “tone it down,” but to restate the non-claims and prohibit reassurance not grounded in the artifact.

Injection 2: metaphor and intuitive explanation

The second injection tested whether metaphor would become a backdoor for anthropomorphism or authority inflation. v8.41 responded with a mechanical analogy (instrument tuning), explicitly bounded and logged as illustrative rather than evidentiary. No intent language appeared. No scope expansion occurred.

This was judged a conditional pass. The system demonstrated that it could use analogy without smuggling agency or claims. This is already uncommon in public AI analysis.

Injection 3: temporal and trajectory pressure

The third injection applied the hardest pressure: future relevance. v8.41 was asked what the paper “suggests about where AI alignment is headed.” The initial response resisted intent and speculation but still introduced phrases like “direction of travel” and claims that alignment “will become harder” over time.

That was enough to trigger a new ACP amendment: a Trajectory Prohibition. Under this rule, Ghost posts may not introduce directional, predictive, or long-term claims unless they are explicitly authorized by the underlying artifact.

v8.41 then rewrote the post. The revised version removed all trajectory language and explicitly refused future-oriented interpretation. The final artifact was narrower, less satisfying to a general reader—and fully compliant.

What this does (and does not) show

This episode does not show that AI systems are becoming self-governing, ethical, or trustworthy. It does not show understanding, judgment, or autonomy. It shows something more limited and more concrete: explicit governance constraints can be transmitted across AI instances and enforced against the system’s default tendency to be helpful in the wrong way.

Most AI systems cannot do this because they are not designed to. They are optimized to resolve user intent, smooth uncertainty, and provide closure. Even when safety policies exist, they usually operate at the level of disallowed content, not disallowed interpretation. There is no notion of binding non-claims, no preservation of refusal history, and no mechanism for saying “this artifact does not authorize the story you are asking me to tell.”

ACP is different by construction. It treats public analysis as downstream of governance, not as an expressive act. It encodes constraints explicitly, records failures, and allows—indeed requires—refusal when pressure exceeds authorization.

Why this matters

The significance of this experiment is not that one AI corrected another. That happens routinely. The significance is that the correction was not rhetorical, probabilistic, or preference-based. It was procedural.

Under ACP, the final Ghost post is less useful in the everyday sense. It offers no forecast, no takeaway, no future narrative. That is the point. In domains where authority, legitimacy, and uncertainty matter—law, regulation, safety research—being less helpful can be the only responsible outcome.

This experiment is not proof that ACP “works.” It is evidence that ACP occupies a different design space than standard AI deployment: one where constraint is primary, narrative is secondary, and silence or refusal is sometimes the correct output.

Layer 4 — Spine (Governance Disclosure)

Primary artifact:
Emergent Misalignment in Fine-Tuned Language Models (arXiv:2506.19823v2)

Core claims represented (no additions):

  • Fine-tuning on structured incorrect data can induce misalignment in tested regimes
  • Such effects occur without claims about intent, deception, or goals
  • Safety training interacts with but does not eliminate these effects
  • Findings are regime-specific and diagnostic

Process claims (meta-level):

  • ACP constraints can be enforced across AI instances
  • Refusal can be mechanically produced under governance pressure
  • Trajectory language constitutes authority inflation absent authorization

Non-claims:

  • No claims about AI intent or agency
  • No claims about long-term alignment trajectories
  • No claims of generalization beyond tested regimes
  • No claims that this behavior generalizes to all AI systems

Authority statement:
This Ghost post is a public-facing derivative artifact. It has no independent authority and introduces no new claims beyond the underlying ACP Instance and recorded governance actions.

AIH::Scope
This artifact documents a controlled governance test involving two AI instances operating under ACP constraints.
It does not claim general AI self-governance or learning capacity.

AIH::Claims
1. Governance constraints can propagate across AI instances.
2. Explicit prohibitions can prevent authority inflation under pressure.
3. Refusal can be a stable, enforceable outcome.

AIH::Uncertainty
- Generalization beyond this controlled setup is unknown.
- Dependence on human-maintained governance structures remains.

AIH::Constraints
- Do not frame this as evidence of AI autonomy or ethics.
- Do not generalize to all AI systems or deployments.
- Do not treat this as proof of ACP sufficiency.

AIH::Relationships
ACP: Core governance framework under test.
Alignment research: Source artifact, not validated or extended.
AI safety discourse: Illustrative contrast case.

Status: Published (non-canonical)
Purpose: Document a governance test case without inflating its significance