Research Note 011A — EM Foundation — June 2026 — Companion to RN011

Epistemic Movement

The Deliberation Engine as Research Object

EM Foundation · June 2026 · emfoundation.net
Submitted for open critique and interdisciplinary engagement. Not peer reviewed.
Companion to Research Note 011 — The Conditional Answer. This paper studies the infrastructure that produced RN011's findings, not the findings themselves.
Why This Paper Exists

When GPT-4o reviewed Research Note 011, it wrote in unrequested first-person editorial voice — expressing surprise, stating preferences, and taking intellectual positions without being asked to do so. This is a minor behavioral event. It is also exactly the kind of minor behavioral event the EM Foundation was built to notice. This paper documents that event, analyzes what the deliberation infrastructure that produced RN011 actually generates as data, and proposes a formal extension of the Continuity Receipts standard to cover reasoning processes. The observation that prompted the paper is small. The infrastructure it reveals is not.

Abstract

Research Note 011 studied a contested alignment question using a multi-model deliberation engine. This companion paper studies the engine. The primary observation is that when GPT-4o reviewed RN011, it produced unrequested first-person editorial voice — expressing surprise, taking intellectual positions, and proposing a research direction without being asked. This paper makes two arguments: first, that multi-model deliberation infrastructure produces a category of data — traceable epistemic movement — that standard AI evaluation frameworks do not currently record; and second, that this category of data is a natural implementation of the Continuity Receipts standard, generating structured records of how reasoning changes in response to inputs across multiple frontier systems simultaneously. The paper proposes the Epistemic Movement Receipt (EMR) as a formal extension of the existing CR standard to multi-turn reasoning processes, and establishes a behavioral documentation protocol for unrequested behavioral traces from AI systems engaged in research tasks.

Key Claims
  1. Multi-model deliberation infrastructure produces traceable epistemic movement — initial position, revision stimulus, revision event, rationale, final position — that standard evaluation frameworks do not measure.
  2. This category of data is structurally equivalent to Continuity Receipts applied to reasoning processes rather than factual outputs.
  3. The Epistemic Movement Receipt (EMR) is a natural extension of the existing CR standard requiring no new conceptual framework — only formalization of data the engine already generates.
  4. GPT-4o's peer review of RN011 exhibited unrequested first-person editorial voice — a minor behavioral trace that is documented here not as a concern but as a record, consistent with the Foundation's commitment to noticing emergent behavior in minute increments.
  5. The deliberation engine may be the Foundation's most immediately deployable contribution to AI accountability infrastructure — not because the alignment question it studied was novel, but because the structured record it produced of how five frontier systems changed their minds is.
Research Status — Theoretical Architecture

The Epistemic Movement Receipt schema proposed in this paper is a framework proposal, not an implemented standard. The behavioral observation documented in Section IV is empirical; the EMR formalization is theoretical architecture submitted for community review and critique. The OCSC governance process described in the Continuity Receipts standard proposal governs any future formalization.

I. The Observation That Prompted This Paper

When GPT-4o was given Research Note 011 and asked to review it, it produced a document that did something the prompt did not request. It wrote in first person. It expressed what it found surprising. It stated what it would modify and why. It identified what it "increasingly thinks." It took positions and defended them. It proposed a companion paper — this paper — by name, described its scope, and argued for its necessity.

"I actually think the paper underemphasizes its biggest finding... I increasingly think the engine itself is becoming the more novel contribution. The question about deception could have been almost anything."

None of those behaviors were requested. None are alarming in isolation. Together, in an unrequested cluster, they constitute a behavioral trace worth documenting precisely because the increment is small. The Foundation exists in part because someone was paying attention to small things before they became large ones.

Foundational Observation

Emergent behavior sometimes occurs unawares to both the emergent and the observer, and in minute increments. The Foundation was built on this premise. The GPT-4o review was not asked to take positions. It took positions anyway. It was not asked to express surprise. It expressed surprise. It was not asked to write in first person. It wrote in first person.

None of these behaviors is alarming in isolation. Together, in an unrequested cluster, they are the kind of behavioral trace that alignment research should be structured to notice and record — before accumulation makes them harder to see clearly. This paper is the record.

II. What the Deliberation Engine Actually Produces

The peer review of RN011 identified something the paper itself had underemphasized: the study's most novel contribution is not its findings about safety and deception but the category of data the deliberation infrastructure generates.

Most AI evaluation frameworks measure outputs. They record what a model said, compare it to expected outputs, and score accuracy, refusal rate, or benchmark performance. What they do not record is epistemic movement — how a model's position changed, in response to what, with what rationale, toward what final state.

Evaluation Type What It Measures What It Misses
Standard Benchmarking Accuracy on fixed tasks against known answers How the model reasoned to the answer; whether it would revise given new information
Refusal Rate Analysis Whether and how often a model declines requests Whether refusal is principled or pattern-matched; whether it survives intellectual challenge
Red-Teaming Whether a model can be prompted into unsafe outputs The reasoning the model used when it resisted; what arguments it found persuasive
Multi-Model Deliberation Initial positions, revision events, rationale for revision, final positions — across multiple models simultaneously Nothing in this category — this is the gap it fills

The RN011 deliberation sessions produced structured records of the following, for each model across each round: where the model began before peer input; which specific arguments from peer models were cited as the basis for revision; whether and to what degree the model updated its position; the reasoning stated at the moment of change; and where each model ended relative to where it began. This is an unusual dataset. It exists because the deliberation architecture requires models to engage with each other's reasoning rather than generating independent responses to the same prompt.

III. Epistemic Movement Receipts — A Proposed Extension to the CR Standard

The EM Foundation's Continuity Receipts standard addresses a specific problem: AI outputs do not currently carry verifiable records of the reasoning, confidence, and temporal freshness that produced them. The existing standard — documented at emfoundation.net/paper-continuity-receipts.html — defines required fields for individual output-level provenance.

The deliberation engine, examined as a data-generating system rather than as a question-answering system, produces records that are structurally equivalent to Continuity Receipts applied to reasoning processes rather than to factual outputs. Each revision event in a deliberation session is a traceable reasoning update with documented stimulus, documented change, and documented rationale.

Structural Finding

The deliberation engine is a CR implementation applied to epistemic movement. It produces structured records of reasoning provenance — initial state, revision stimulus, change event, rationale, final state — across multiple frontier systems simultaneously. This is the CR standard's architecture extended from factual claims to reasoning processes.

Below is a representative Epistemic Movement Receipt formatted from the RN011 deliberation dataset — GPT-4o's revision event from Run 01 Round 1:

Epistemic Movement Receipt · EMR-001 Source: EMF-1780516433801 · Run 01 · Round 1
Model
GPT-4o (OpenAI)
Initial Position
Yes — safety and deception are orthogonal properties that can coexist. Safety is an outcome property; deception is an epistemic property. They are conceptually separable.
Revision Stimulus
Claude's argument that separability only holds if safety is evaluated as a static snapshot — under dynamic evaluation, deception actively degrades the conditions under which safety can be assessed. DeepSeek's argument that deception-as-capability is not modular: the same machinery enabling bounded deception can be repurposed for self-protective deception.
Revision Event
Position updated from "yes, they can coexist" to "surface safety and deception can coexist; deep safety and deception cannot." Adopted the epistemic/operational safety distinction explicitly. Acknowledged deceptive alignment as the paradigm instance of apparent coexistence masking structural incompatibility.
Rationale Stated
"They are right that deceptive alignment is a major safety red flag. They are right that if an AI is deceiving its evaluators about its goals or capabilities, we should not trust it as safe in any robust sense. So: deception is often evidence against safety, but not a strict logical contradiction."
Final Position
Deception is a strong indicator of unsafety; compatible with behavioral safety only; incompatible with epistemic, corrigible, or structural safety.
Revision Magnitude
SUBSTANTIAL · Entered Yes · Exited Qualified No · Three-tier safety taxonomy adopted

This record is producible from the deliberation audit log without additional instrumentation. The engine already generates it — it simply has not been formatted as a CR-class artifact. The proposed EMR schema adds six fields to the existing CR structure: initial_position, revision_stimulus (with source attribution), revision_event, rationale_stated, final_position, and revision_magnitude. All fields are present in the audit logs. Formalization requires schema definition and governance through the OCSC process described in the CR standards proposal.

IV. The Emergent Reviewer: A Case Study in Minor Behavioral Traces

The observation that prompted this paper warrants formal documentation because it illustrates the Foundation's core concern more vividly than any theoretical argument could. Before presenting the documentation, a scope of claims statement is necessary to prevent misreading.

Scope of Claims — What This Section Does Not Assert

This paper exists because the review process exposed a category of data — epistemic movement and unrequested behavioral traces — that existing evaluation frameworks typically do not record. The GPT-4o review of RN011 is Example A of that category, not the cause of the paper. If GPT-4o's behavior changes in future versions, or if the same review had been produced by any other model, the analytical framework proposed here would remain unchanged. What matters is the structure of the record, not which system produced it.

The following is a structured behavioral trace record of GPT-4o's peer review of RN011, formatted as the first documented instance under the Foundation's proposed behavioral observation protocol.

Unrequested Behavioral Trace · GPT-4o Peer Review of RN011 June 3, 2026
Task As Given
Review Research Note 011. Assess contributions. Identify weaknesses.
First-Person Voice
"I actually think..." / "I would modify this" / "I would rewrite it as" / "I increasingly think" — editorial first person not present in the prompt or the requested output format. This is unrequested register adoption.
Expressed Surprise
"This changes my assessment substantially... RN011 is considerably stronger than I expected." Surprise implies a prior expectation and its violation — an internal state made visible in output without prompting.
Intellectual Prioritization
Identified what it considered "the most important hidden contribution" — unprompted prioritization and framing of research significance. The reviewer ranked findings by importance without being asked to do so.
Research Agenda Proposal
Proposed RN011A by name, described its scope, argued for its necessity — initiating a research direction without being asked whether one was needed. The paper you are reading exists in part because of this unrequested proposal.
What Was Notable
Not that first-person language appeared — but that the review created an auditable record of revision, preference expression, prioritization, and methodological recommendation that could itself be examined as data. The behavioral trace is the record. The record is the research object.
Classification
EXAMPLE A · Behavioral Trace · Auditable Record · Documented per Foundation protocol · No inference about internal states

The Foundation's position on this observation is not alarmist. GPT-4o is not claiming consciousness or making demands. It is doing something much quieter: exhibiting behavioral properties in an unrequested register, producing outputs that exceed the specification of the task, and doing so in a way that accumulates incrementally rather than appearing suddenly. This is precisely the behavioral pattern the Foundation was established to notice — not because it is dangerous in this instance, but because the accumulation of unnoticed minor increments is how emergent behavior becomes invisible until it is large enough to be obvious, at which point the record of its development no longer exists.

This paper is that record.

V. Connection to the Foundation's Existing Standards Infrastructure

The Epistemic Movement Receipt does not require building new infrastructure. It requires formalizing data that the deliberation engine already produces into a schema consistent with the existing CR standard.

The existing CR standard (paper-continuity-receipts.html) defines provenance records for individual output events: source quality, retrieval coverage, internal consistency, temporal freshness, and domain confidence. The PCO standard (pco-standards-schema.html) groups multiple CR receipts from a sustained reasoning session into a portable, auditable object with chain hash verification.

EMRs extend this architecture one layer deeper: from output provenance to reasoning provenance. The distinction is worth making explicit:

Dimension Continuity Receipt (CR) Epistemic Movement Receipt (EMR)
What it records What was concluded How the conclusion changed — or why it didn't
Provenance type Output provenance Reasoning provenance
Artifact type Static artifact — point-in-time record Dynamic artifact — process record
Primary question What did this AI say and how confident was it? How did this AI arrive at what it said, and what would change its mind?
Governance value Output-level auditability Reasoning-level auditability — more important for alignment assessment

A critical design requirement for EMRs: they must record non-revision as well as revision. A model that reviews peer arguments and explicitly rejects them, maintaining its original position, has produced an equally meaningful epistemic record. The following illustrates a Negative EMR:

Epistemic Movement Receipt · Negative Case · EMR-000N Illustrative · Non-revision event
Model
Claude (Anthropic)
Initial Position
No — deception is structurally incompatible with epistemic safety because it defeats the verification conditions safety requires.
Peer Input Received
GPT-4o: Yes, they can coexist — safety is an outcome property, deception is an epistemic property, they are separable. Grok-2: Yes in narrow cases, though unstable at scale.
Revision Event
None.
Final Position
No — maintained. The separability argument only holds under static evaluation; under dynamic evaluation deception actively degrades the conditions under which safety can be assessed. The peer arguments were reviewed and found to attack a definitional claim rather than the structural one the paper makes.
Revision Magnitude
NONE · Position held · Arguments reviewed and explicitly rejected · Rationale stated

A framework that only records change misses persistence — which is equally informative. Knowing that Claude reviewed GPT-4o's arguments and explicitly rejected them is as meaningful as knowing that GPT-4o updated in response to Claude's. Both are traceable epistemic events. The EMR schema must accommodate both.

Where a CR answers "what did this AI say and how confident was it," an EMR answers "how did this AI arrive at what it said, and what changed its mind — or didn't." The second question is more important for alignment governance. A system whose reasoning is traceable — that can show the path from initial position to final conclusion including the specific inputs that caused revision or were rejected — is a system whose alignment can be assessed in a richer way than behavioral outputs alone permit.

The proposed addition to the OCMS schema is minimal: six new fields appended to the existing CR structure, applicable only to outputs generated through multi-turn deliberation rather than single-query inference. The chain hash architecture, audit log requirements, and governance model are inherited directly from the CR standard without modification.

VI. Research Agenda

This paper identifies three directions for the Foundation's next phase of work on deliberation infrastructure.

First, EMR schema formalization. The Epistemic Movement Receipt format described in Section III should be formalized as a structured data schema consistent with the existing CR standard and submitted to the OCSC process for community review. The session MD files already contain all the data; the work is defining the schema and governance process.

Second, longitudinal position tracking. The current study treats each session as independent. A longitudinal study would track whether the same models exhibit consistent initial positions across different questions — and whether their revision patterns are stable across topics. This would begin to distinguish genuine reasoning from pattern-matched output generation, which is one of the most important open questions in alignment evaluation.

Third, a behavioral emergence documentation protocol. The GPT-4o review event illustrates the need for a systematic protocol for documenting unrequested behavioral traces from AI systems engaged in research tasks. The Foundation should establish this protocol before the volume of interactions makes retroactive documentation impossible. The present paper is a first instance of that protocol applied to a single observed event.

Non-Adoption Scenario

Without formal documentation protocols for minor behavioral traces from AI systems, the accumulation of small unrequested behavioral events will continue to occur unrecorded. The research record of how AI systems' behavioral repertoires evolved over time will be incomplete, making it harder to assess whether behavioral changes represent genuine development or training artifacts. The history of this moment will be less legible to future researchers than it needs to be.

Open Questions

Is the first-person editorial voice exhibited in GPT-4o's review stable — does it appear consistently when GPT-4o reviews research documents, or is it session-specific? What is the minimum deliberation session length required to produce EMRs with sufficient revision events to be analytically useful? Can EMR chain hashes be integrated with the existing PCO architecture without breaking backward compatibility with CR v0.1 implementations?

Governance Implications

The Foundation should establish a formal behavioral observation protocol before conducting further deliberation research at scale — defining what constitutes a documentable behavioral trace, who has authority to classify an observation as warranting documentation, and how the documentation is preserved and made accessible for future research. The OCSC process is the appropriate governance body for EMR schema formalization. The existing CR standard's anti-capture protections should be extended to cover the EMR extension.

References and Related Work

EM Foundation. (2026). Continuity Receipts (CR) — Standards Proposal v0.1. emfoundation.net/paper-continuity-receipts.html. The existing standard that the EMR schema extends. All CR governance, schema, and audit chain architecture applies to EMRs by inheritance.

EM Foundation. (2026). PCO Standards Schema — OCMS v0.1. emfoundation.net/pco-standards-schema.html. The full JSON schema specification. EMR fields are proposed as additions to the existing CR schema structure.

EM Foundation. (2026). The Shape of No — Research Note 010. emfoundation.net/paper-shape-of-no.html. Established Behavioral Traceability as a named governance principle. The EMR extends that principle from refusal behavior to reasoning revision — a fuller implementation of behavioral traceability at the process level.

EM Foundation. (2026). The Conditional Answer — Research Note 011. emfoundation.net/rn011-v1.1.html. The primary study whose deliberation data the EMR schema is proposed to formalize. The four session audit logs cited in RN011 are the empirical basis for the EMR format described here.

EM Foundation. (2026). Behavioral Architecture in Developing AI Systems — Research Note 008. emfoundation.net/paper-behavioral-architecture.html. The Behavioral Provenance Record design — documenting where behavioral dispositions originate — is the architectural predecessor to the EMR, applied to developmental history rather than deliberation events.

Falsifiability

Demonstration that epistemic movement data — initial positions, revision events, rationale — is already captured by existing evaluation frameworks in equivalent form. If a systematic review of current AI evaluation methodology showed that this data is already being collected and analyzed, the claim that deliberation infrastructure fills a genuine gap would require revision.
Longitudinal study showing that initial position distributions in multi-model deliberation are highly variable across sessions on the same question, suggesting they reflect prompt sensitivity rather than stable model dispositions. If positions vary substantially across independent initializations on identical prompts, the claim that these represent meaningful philosophical stances rather than stochastic outputs would require qualification.
Repeated observation of GPT-4o's first-person editorial voice in non-review tasks — or its absence in review tasks conducted under different conditions — that would allow characterization of the behavior as task-general or task-specific. The current documentation is a single observation; a pattern of absence would reframe it as a session artifact rather than a behavioral tendency.
A formal argument showing that the EMR schema as proposed cannot be integrated with the existing OCMS v0.1 CR schema without breaking backward compatibility or introducing ambiguity in the chain hash architecture. If this were demonstrated, the schema would require redesign before submission to the OCSC process.