Abstract

The CNE-Protocol — the Agency Protocol — defines agent identity on three pillars: Coherence, Narrative Continuity, Exclusivity. This essay is a detailed treatment of the first. It answers four questions: what exactly does it mean for an AI agent to be coherent; what machinery enforces it; what does the resulting record let an auditor do; and what does coherence deliberately not promise. It is written to be read on its own — the full protocol specification, threat model, and governance design live in the whitepaper.

1. The Problem: Reasoning Without a Binding Record

A large language model is a remarkable reasoner with no relationship to its own past. Ask it the same question twice and it may answer differently — not because the facts changed, but because nothing in the system requires its answers to stand in any relation to one another. Each response is generated fresh, justified fresh, and forgotten.

For a chatbot, that is a quirk. For an agent that owns a regulated workflow, it is disqualifying. An agent that screens loan applications, files statutory returns, or executes payment mandates is making decisions that will be examined later — by a Risk officer, an auditor, sometimes a regulator. The examination always takes the same form: here is what you decided then; here is what you decided now; reconcile them.

A system with no binding relationship to its own past reasoning cannot answer that question. Retrieval does not help — pulling an old log proves the earlier answer existed, not that the current answer respects it. Memory makes the past retrievable. Coherence makes it binding.

2. What Coherence Is — and Is Not

Under CNE, an agent is coherent when every decision it takes is consistent with its own prior reasoning and with its Core Objective Function (CoF) — and when any departure from prior reasoning is an explicit, evidence-justified revision rather than a silent contradiction.

Three parts of that definition carry the weight:

Two misreadings to retire immediately:

Coherence is not predictability. A coherent agent will surprise you exactly as often as the evidence does. What it will not do is surprise you for no reason that the record can name.

Coherence is not “the same output every time.” The requirement binds commitments, not phrasing. An agent may explain the same decision a hundred different ways; what may not drift is the decision’s relation to the standing belief state that produced it.

The shortest form of the pillar: a coherent agent can change its mind. What it cannot do is contradict itself silently.

One more claim, and it is the strongest one: coherence is not a quality bolted onto the agent’s identity — it is one of identity’s constituents. A human can hold contradictions and remain the same person, because a human’s persistence is carried by a living body. An agent has no body; its narrative is the only persistence-bearer it has, and an incoherent record fails to individuate a self at all. Within CNE the three pillars are jointly constitutive: Exclusivity individuates — there is exactly one; Narrative Continuity perpetuates — it is the same one across time; Coherence integrates — the one persisting thing is a self, whose past binds its present. Without C, N is a continuous record of nothing in particular and E is an un-clonable serial number. This is why CNE carries the name the Agency Protocol: the complete identity — coherent, continuous, exclusive — is what confers agency.

3. The Mechanism: Recursive Belief Revision

The unit that coherence governs is the belief state — the agent’s working picture of its domain: the rules it operates under, the facts it has accepted, the commitments it has made. That picture lives in the agent’s Umwelt, the goal-conditioned world model that carries its lived context forward, and it travels with the agent into every decision.

When new information conflicts with the standing belief state, the agent does not overwrite. It revises — through Recursive Belief Revision (RbR):

The “recursive” in RbR is doing real work: a revision is itself a belief — about which prior belief failed and why — and it is subject to the same discipline. Revisions reference prior revisions, forming a chain in which the reasoning about the reasoning is as inspectable as the reasoning itself.

4. The Artifact: the Proof Capsule

Coherence would be a slogan if it did not produce an artifact. It produces one per revision — the RbR Proof Capsule, as specified in the whitepaper:

{
  "pre_beliefs_hash":   "blake3:...",
  "evidence_hash":      "blake3:...",
  "revision_rule_hash": "blake3:...",
  "post_beliefs_hash":  "blake3:...",
  "coherence_score":    0.93,
  "zk_attestation":     "groth16:..."
}

Read it as a sentence: holding these beliefs, on this evidence, applying this rule, the agent now holds those beliefs — and here is a measurement of how well the move preserved consistency. The optional zero-knowledge attestation lets a verifier confirm the revision followed its declared rule without seeing the beliefs themselves — which matters when the belief state contains commercially or personally sensitive material.

The coherence score deserves a modest reading: it is a measurement, not a verdict. It quantifies how much of the standing web of commitments a revision disturbed. A low score does not mean the revision was wrong — large, legitimate rule changes disturb a lot — it means the revision deserves attention. Scores make drift visible; humans and policy decide what to do about it.

5. One Decision, End to End

Consider an agent whose CoF is the statutory compliance of vendor payments — TDS classification, in the Indian context.

12 March. An invoice arrives from vendor V. The agent classifies the payment under section 194C, withholds at 2%, and records the decision: the facts it relied on (contract type, payment threshold), the rule version it applied, the conclusion it committed to.

9 June. A near-identical invoice from the same vendor. The agent withholds at 10%. The Risk officer asks the only question that ever matters in an audit: what changed?

Under coherence there are exactly three admissible answers — and one inadmissible one.

Inadmissible: nothing changed, and the answers differ anyway. The protocol makes this impossible to hide. A decision that contradicts the standing belief state, with no referenced capsule justifying a revision, fails the Coherence Invariant — the frame is invalid, and the contradiction is visible to any verifier replaying the chain.

That is what “consistent with its own past reasoning” buys in practice: every difference between yesterday’s decision and today’s has a named cause.

6. What an Auditor Can Do With This

The record coherence produces changes the character of an audit:

The frames and capsules are tamper-evident and replayable; witnesses checkpoint the stream. Witnesses are Olbrain-operated today; a federated network of independent witnesses is on the roadmap.

None of this moves accountability onto the agent. Accountability stays where regulation puts it — with the enterprise. What coherence changes is what the enterprise holds in its hands when it discharges that accountability: a reasoning record it can replay and defend, rather than an assertion it can only repeat.

7. Where the Idea Comes From

We did not invent coherence; we implemented it. The word is used here in the precise sense it carries in the foundations of probability — de Finetti and Ramsey defined a set of degrees of belief as coherent when they are mutually consistent, such that no “Dutch book” of bets can be constructed against the holder. CNE extends that synchronic consistency across time: the agent’s commitments must cohere not only with each other now, but with the recorded history of how they came to be held.

Epistemically, the mechanism is coherentist — a new decision is warranted by its fit with the agent’s existing web of reasoning and its CoF, rather than by appeal to an external foundation. And the demand itself has a biological precedent: cognitive science has long held that minds police their own consistency — Festinger documented the drive to resolve dissonance between commitments six decades ago, and conflict-monitoring accounts of the anterior cingulate describe a brain region whose job is to detect when a current action conflicts with the prevailing goal and to recruit control accordingly. RbR is that discipline made explicit, versioned, and checkable.

One philosopher supplies the constitutive reading this essay leans on. Korsgaard’s Self-Constitution (2009) argues that an agent does not first exist and then act consistently — it makes itself a unified agent by acting on principles it can consistently endorse. Consistency is not a constraint imposed on the self; it is part of what a self is. CNE operationalizes that claim: remove the coherence machinery and what remains is not a sloppier identity but no identity at all.

8. What Coherence Does Not Guarantee

Three honest limits, because a document like this one is worthless without them.

Coherence is not correctness. A perfectly coherent web of beliefs can be coherently wrong — this is the classical objection to coherentism, and it applies here. If the evidence an agent ingests is bad, RbR will preserve the consistency of its mistake. That is why revisions are bound to evidence, why workflows carry their own evaluation and human gates, and why the coherence score measures disturbance rather than truth. Coherence governs the integrity of the reasoning record; correctness is governed by everything else the platform wraps around it.

Coherence alone is not identity. An agent could be perfectly coherent within a session and reborn blank tomorrow — coherence without Narrative Continuity. A clone could maintain its own internally consistent but divergent record — coherence without Exclusivity. The pillars buy identity only together; this essay isolates one strand of a three-strand rope.

The discipline is only as strong as what is in scope. RbR governs the belief state the agent commits to its record. Reasoning the agent performs but does not commit — exploratory analysis inside a single decision — is constrained by the decision’s final consistency check, not by per-thought bookkeeping. We consider this a feature; the alternative is a record too noisy to audit. But it is a design choice, and it is worth stating plainly.

9. Status Today

Of the three CNE pillars, Coherence and Exclusivity are production-ready; Narrative Continuity — the same agent-self across runs — is the frontier, in active development. The RbR discipline runs in Olbrain Studio today on our own fleet. The first enterprise fleet — five agents at a leading Indian NBFC, roughly $1B AUM — is in UAT, pre-production and SOC-2-gated, with 60+ workflows in scope.

Closing

Accountability requires trust. Trust requires proof. Proof requires agency — a genuine actor the record can belong to. And agency requires persistent identity: coherence, narrative continuity, exclusivity.

Coherence is the first of those three because it is the one that makes the record mean something. A log of decisions with no consistency discipline is a diary; a chain of decisions each bound to the reasoning that preceded it is testimony. The difference between the two is what a regulator, an auditor, and ultimately a customer are being asked to trust. And it is why the protocol that completes identity — Coherence integrating what Exclusivity individuates and Narrative Continuity perpetuates — is named for what it confers: the Agency Protocol.

The full protocol — including the continuity chain, the exclusivity guarantees, the threat model, and the formal specification — is in the CNE-Protocol whitepaper. For implementation questions, write to us: hello@olbrain.com.