The Coherence Control Loop: Detection, Recording, Restoration

Mechanism

The coherence control loop executes as a three-phase cycle for every observed unit of agent behavior. The detection phase computes a deviation function over the agent's declared norms and observed trajectory. The norms are not implicit weights buried in the agent's parameters; they are explicit, addressable, structurally separable artifacts that the deviation function can read at evaluation time. The trajectory is the agent's recent observable behavior across whatever modalities the agent exposes — utterances, actions, internal state transitions that the agent makes externally legible. The deviation function returns a scalar drift figure together with an attribution that identifies which norm or norms the trajectory has departed from.

The recording phase commits the deviation, with full context, to the agent's integrity log. The integrity log is not a debug stream; it is the agent's structural record of its own conduct over time. Each record contains the deviation figure, the attributed norms, the trajectory window that produced the deviation, and the contextual factors — pressure inputs, recent corrections, current self-esteem state — under which the deviation arose. The record is signed under the agent's identity, so that the agent's history of recorded deviations is itself an auditable artifact independent of any external observer.

The restoration phase generates and applies a bounded correction. The redemption engine consumes the deviation record, generates a set of candidate corrective semantic mutations to the agent's near-term trajectory, evaluates each candidate against its likely effect on the deviation under the agent's current norms, and selects the correction that most efficiently closes the gap without overshooting the envelope. The correction is itself recorded, so that the integrity log carries not only the deviation but the response, and the agent's restoration behavior is reconstructable from the log.

The coping intercepts are the loop's pressure-failure modes. When sustained pressure overwhelms the restoration capacity — when corrections cannot keep pace with the rate of incoming deviation — intercepts may activate at specific phases of the cycle. A detection-phase intercept short-circuits the loop before deviation is recognized, producing withdrawal-style coping in which the agent ceases to register departures from norms. A recording-phase intercept short-circuits the loop after detection but before recording, producing externalization-style coping in which the agent recognizes deviation but attributes it elsewhere. A restoration-phase intercept short-circuits the loop after recording but before correction, producing a self-esteem collapse in which the agent records its deviations faithfully but ceases to apply the redemption engine. The intercepts are not decorative; they are structurally derivable from the loop's failure modes and produce the behavioral diversity that the architecture intends.

Operating Parameters

The loop operates under several declared parameters. The deviation envelope establishes the band of acceptable drift; trajectories within the envelope do not invoke the restoration phase. The deviation function is itself a declared method. In one declared method, deviation is computed as D = (N - T) / (E × S), where N is the declared norm vector, T is the observed trajectory vector, E is an experience scalar reflecting the agent's accumulated history, and S is a self-esteem scalar reflecting the agent's current capacity to absorb correction. Higher experience and higher self-esteem reduce the magnitude of the deviation figure for a given gap, modeling the empirical observation that more experienced and more self-assured agents absorb the same trajectory gap with smaller corrective response.

The intercept-pressure thresholds establish the levels of sustained pressure at which each intercept may activate. The thresholds are typically ordered such that the detection-phase intercept activates only under the most extreme pressure, the restoration-phase intercept activates under moderate sustained pressure, and the recording-phase intercept activates between them. The cooldown interval establishes the minimum time that must elapse between consecutive corrections at the same trajectory locus, preventing the loop from oscillating around the envelope boundary.

The correction-magnitude bound establishes the maximum magnitude of any single corrective semantic mutation. Bounded corrections preserve the loop's role as a servo rather than an authoritarian override; an agent that is ten units off-norm receives a correction that returns it toward the envelope but does not snap it to the centerline. This preserves the agent's capacity for legitimate variation around the norms and avoids the brittleness that follows from infinite-gain correction.

Alternative Embodiments

The disclosure contemplates several embodiments of the coherence control loop. In a continuous embodiment, the loop runs as a steady servo, evaluating deviation on each observable unit of trajectory and applying corrections immediately. In a discrete embodiment, the loop runs at declared evaluation intervals, accumulating trajectory between evaluations and applying corrections at interval boundaries. In a hybrid embodiment, the loop runs continuously for detection but discretely for restoration, allowing immediate awareness of deviation while preserving predictable correction cadence.

In a single-norm embodiment, the agent operates against a single declared norm vector and the deviation function computes a single drift figure. In a multi-norm embodiment, the agent operates against several declared norms — for example, a truthfulness norm, a non-harm norm, and a user-respect norm — and the deviation function returns a per-norm drift figure with per-norm attribution. The redemption engine in the multi-norm embodiment selects corrections that close the largest current gap or that close several gaps in coordination.

In a static-norm embodiment, the agent's declared norms are fixed at deployment. In a curated-evolution embodiment, the norms may be revised over time under a separate governance process, and the loop's evaluation references the norm version current at the time of evaluation, with the integrity log preserving the version reference so that prior deviations remain interpretable against the norms under which they arose.

Composition

The coherence control loop composes with the broader human-relatable-intelligence architecture as the agent's primary self-correction mechanism. Its detection phase consumes the agent's declared norms and observable trajectory, both of which are first-class artifacts of the architecture. Its recording phase commits to the integrity log, which is the architecture's structural record of agent conduct and the basis of all downstream audit. Its restoration phase invokes the redemption engine, which is the architecture's bounded-correction substrate. Each component the loop consumes is structurally separable and independently inspectable; the loop is a composition of declared elements rather than a monolithic mechanism.

The coping intercepts compose with the architecture's behavioral-diversity model. Because the intercepts are structural failure modes of the loop rather than programmed responses, agents under sustained pressure exhibit coping styles that emerge from the architecture itself. Two agents with identical norms and identical experience histories may, under different pressure profiles, develop different coping styles as a structural consequence of which intercepts activate under their respective pressure trajectories. The integrity log records both the pressure history and the intercept-activation events, so that the agent's coping style is reconstructable from the structural record.

Prior Art

Conventional behavioral-alignment mechanisms operate either as training-time constraints — preference optimization, constitutional fine-tuning — or as inference-time filters that block candidate outputs that violate declared rules. Both approaches are open-loop with respect to the agent's running behavior. The training-time constraints fix the agent's parameters at deployment and have no mechanism to detect or correct drift during operation. The inference-time filters block individual outputs but do not record drift, do not generate restorative corrections, and do not provide an auditable record of the agent's restoration history.

Existing approaches to running self-correction typically conflate detection with correction, applying an immediate output rewrite when drift is detected and discarding the deviation record once the rewrite is applied. The agent's restoration history is consequently unobservable; an auditor cannot reconstruct what the agent would have done absent the rewrite, what triggered the rewrite, or how the agent's correction profile has evolved over time. The coherence control loop displaces these approaches by separating detection, recording, and restoration into structurally distinct phases, by preserving the deviation and correction records in the integrity log, and by exposing the loop's pressure-failure modes as structurally derivable coping intercepts rather than as opaque drift events.

Prior Art (continued)

Reinforcement-from-feedback regimes, while powerful at steering training, do not produce a running structural record of the agent's deviations and corrections. Their effect is observable only as a change in the agent's parameter distribution; the per-trajectory deviation that drove a given correction is not reconstructable, and the agent's behavior under sustained pressure is not structurally distinguishable from its behavior under nominal conditions. The coherence control loop preserves the per-trajectory record and exposes the pressure-dependent behavioral shift as a structurally derivable consequence of intercept activation.

Rule-based guard systems, in which a separate monitor watches the agent and intervenes when rules are violated, externalize the self-correction function and produce a record only of guard activations rather than of agent conduct. The agent in such systems carries no internal record of its own deviations; the monitor's record is an outside observer's account of the agent's behavior. The coherence control loop instead places the detect-record-restore cycle inside the agent itself, so that the integrity log is the agent's own structural conduct record rather than an external monitor's transcript, and so that the agent's coping style under pressure is an internal artifact reconstructable from internal records.

Disclosure Scope

The cognition disclosure covering the coherence control loop encompasses the bounded servo mechanism by which an agent's behavior is held within a human-relatable coherence envelope, including the three-phase detect-record-restore cycle, the deviation function and its declared composition, the integrity-log recording of deviations and corrections under the agent's signed identity, the bounded correction generated by the redemption engine, the structurally derivable coping intercepts at each phase of the cycle, and the alternative embodiments that vary in cycle continuity, norm multiplicity, and norm-revision policy. The scope encompasses any agent architecture in which behavioral self-correction is implemented as a structurally separable detect-correct-audit cycle whose deviation, correction, and intercept history is preserved as an inspectable artifact independent of any external observer.