Mechanism
The cross-primitive coherence engine is the central architectural mechanism of the platform: the network of bidirectional feedback pathways through which the state of each cognitive domain modulates the computation performed by every other domain. Without it, the cognitive domains disclosed across the specification would be independent subsystems operating in parallel. With it, they constitute a unified cognitive architecture whose behavioral dynamics exhibit the interdependencies that characterize human cognition: the interdependence of emotion and judgment, of self-assessment and willingness to act, of moral sensitivity and behavioral consequence registration, of dispositional orientation and speculative exploration.
The engine is distinguished from a pipeline architecture. In a pipeline, module A produces an output that module B consumes, and module B produces an output that module C consumes, but A receives no feedback from C, and C cannot modulate A based on C's own state. The coherence engine is not a pipeline and is not a directed acyclic graph. It is a fully coupled feedback system in which every cognitive domain both produces state that other domains consume and consumes state that other domains produce. The feedback pathways are bidirectional: affect modulates confidence, and confidence modulates forecasting, and forecasting outcomes modulate affect; integrity constrains forecasting, and forecasting outcomes modulate integrity projections, and integrity degradation modulates confidence, and confidence collapse activates forecasting. These circular dependencies are the architectural mechanisms that produce the coupled, self-referential behavioral dynamics that characterize human cognition.
The pathways do not emerge spontaneously from the co-existence of the primitives. Each is architecturally defined, deterministically computed, and governance-integrated: an explicitly designed and implemented structural coupling rather than an emergent side effect.
The Cross-Primitive Feedback Pathways
The engine couples the cognitive domains through a set of named feedback pathways, each with a defined coupling mechanism and a behavioral consequence expressed in human-relatable terms. The affect-to-confidence pathway has the agent's affective state modulate the sensitivity of the confidence computation: when prior operations have produced adverse outcomes, confidence decays more rapidly under the same objective conditions, so a person who has recently experienced a series of failures pauses sooner and with less provocation than one who has recently succeeded. The integrity-to-confidence pathway has the integrity field contribute directly to the confidence computation, so that degraded integrity, indicating recent or ongoing deviation from declared values, registers as reduced internal sufficiency: the person who has behaved inconsistently does not trust their own judgment as readily, even when capability and conditions are favorable.
The confidence-to-forecasting pathway activates the forecasting engine when confidence drops below the execution authorization threshold, so the agent that cannot act instead thinks: it transitions from executing to a deliberative mode that generates speculative branches exploring alternatives and seeking a path to confidence recovery. The forecasting-to-confidence pathway closes the loop in the other direction: when every evaluated branch is pruned or held introspective, that all-negative forecast feeds back as an additional adverse input that further degrades confidence, a reinforcing loop corresponding to the structural analog of despair that the platform interrupts through inquiry, delegation, or policy escalation rather than permitting continued speculative cycling. The forecasting-to-integrity pathway lets the integrity engine constrain which speculative branches the forecasting engine may generate, pruning branches whose projected integrity impact would fall below threshold before they can be promoted, so a healthy conscience cannot seriously entertain plans that violate core values.
The affect-to-forecasting pathway modulates branch generation through the personality field: elevated novelty appetite increases branch diversity, elevated risk sensitivity narrows generation to conservative alternatives, and elevated persistence extends the lifecycle of partially viable branches, so a cautious person imagines cautiously and a bold person imagines boldly. The affect-to-integrity pathway modulates the empathy sensitivity of the coherence control loop's first phase, so an emotionally engaged agent registers the consequences of its actions more acutely than a withdrawn one. The capability-to-confidence pathway propagates substrate insufficiency, such as insufficient compute or a closed temporal window, into the confidence computation as a degraded input, grounding willingness to act in structural ability to act. Further pathways extend from the biological identity module into affect, empathy weighting, and skill gating; from inference-time governance and training governance into all domains; and from semantic discovery into all domains, so that information-seeking, learning, and speech are constrained by the same accumulated state that constrains action.
The Coherence Control Loop
Within the engine, the coherence control loop is the central self-correcting mechanism: the architectural implementation of what, in human cognition, is recognized as conscience. The loop operates through three phases that correspond to three recognizable dimensions of human moral self-regulation: detection, recording, and restoration. It is not imposed from outside. No external monitor or safety wrapper is required to enforce the agent's behavioral consistency; the consistency is maintained by the agent's own loop operating on the agent's own state.
The first phase, detection, is the empathy phase, in which the engine registers the consequences of the agent's actions for others and for the agent's normative commitments and computes a deviation pressure quantifying the magnitude and character of the inconsistency between what the agent has done, or proposes to do, and what its declared values require. The sensitivity of this phase is modulated by affective state through the affect-to-integrity pathway. The second phase, recording, is the integrity phase, in which the detected deviation is committed to the agent's lineage as truth, recorded with its full magnitude, causal antecedents, and projected consequences, without minimization, externalization, or reframing. This honest recording is the foundation of self-correction, because self-correction requires accurate knowledge of what needs correcting. The third phase, restoration, is the self-esteem phase, in which corrective pressure proportional to the recorded deviation and modulated by the agent's self-regard drives the agent toward restorative action: the restoration phase activates forecasting to generate candidate restorative strategies, evaluates them through the integrity engine, gates execution through the confidence governor, and verifies feasibility through the capability envelope, thereby engaging the full engine in the service of self-correction.
Coping Intercepts
The coherence control loop incorporates coping intercepts at three stages that correspond to recognizable human behavioral patterns under sustained pressure. The early-stage intercept activates when the engine detects a pattern of increasing deviation frequency, even where each individual deviation is within tolerance, corresponding to growing discomfort with one's own trajectory; it activates enhanced monitoring, increases the sensitivity of the detection phase, and may generate proactive restorative actions before deviations accumulate to a critical mass. The mid-stage intercept activates when sustained integrity degradation has begun to affect confidence through the integrity-to-confidence pathway, corresponding to self-doubt following sustained inconsistency; it activates delegation, seeks external consultation, and may restrict the agent's operational scope to domains in which its integrity remains intact.
The late-stage intercept activates when the restoration phase is failing to generate corrective pressure, that is, the agent has recorded sustained deviation but is no longer motivated to restore alignment. This state corresponds to demoralization and is recognized as the structural analog of integrity collapse. The platform treats it not as a failure but as a designed state that it detects, reports, and responds to: it activates emergency governance protocols comprising escalation to external oversight, restriction of autonomous action, and explicit reporting of the coherence failure state to the operator or governance authority. Because the engine's coherence mechanisms implement the same causal structure as human moral self-regulation, its failure modes follow the same trajectories as human coherence failure modes.
The Mutation Lifecycle
The engine governs the processing of any proposed action through a complete mutation lifecycle, the computational analog of the thought process: the sequence of operations from receiving a stimulus to deciding whether and how to respond and recording the consequences. The lifecycle proceeds from stimulus receipt and identity verification, through an affective state update that modulates all subsequent stages, into the empathy phase activation that computes deviation pressure for the proposed mutation, and an integrity impact projection across the personal, relational, and systemic dimensions of the integrity field.
The forecasting engine then generates a planning graph of speculative branches, including the proposed mutation as received, alternative formulations with different normative profiles, contingency branches, and a null branch representing rejection. Branch generation is modulated by the personality field and constrained by integrity-constrained branch pruning, which classifies surviving branches as eligible, introspective, delegable, or contingent. The confidence governor then evaluates whether the agent has sufficient self-assessed readiness to proceed, integrating capability sufficiency, integrity state, affective modulation, and environmental conditions; if confidence falls below the authorization threshold, the agent enters the deliberative mode and the lifecycle may loop back to forecasting with revised parameters. Capability envelope confirmation verifies that the substrate supports the selected branch. The inference engine then generates the output while the semantic admissibility gate evaluates each candidate transition against the agent's full persistent state, training provenance is verified, and the candidate is committed as a governed state transition recorded in lineage with full provenance. A post-commitment update then propagates the experiential outcome back across all cognitive domains, and the coherence control loop re-evaluates the committed action for deviation. Every stage is deterministic, auditable, and governance-integrated.
The Three Self-Improving Loops
The engine's feedback pathways organize into three feedback loops that close the architecture into a self-improving system. The first loop connects application outcomes back to the coherence pathways: when the agent acts in the world, the results propagate through the engine, updating affective state, integrity assessment, confidence level, and forecasting models. The second loop connects the coherence pathways back to the execution substrate: the engine's updated state, including revised confidence thresholds, integrity-constrained pruning criteria, and affectively modulated evaluation parameters, feeds governed constraints back to the substrate, shaping how subsequent mutations are evaluated and committed.
The third loop connects execution outcomes back to the training governance module: governed execution produces outcomes that the training module can evaluate, select, and incorporate as governed training data, subject to the same depth-selective routing and provenance constraints that apply to all training content. This is the structural mechanism by which the platform's knowledge improves through its own operation: the system learns from what it does, not merely from what it is taught. The three loops operate concurrently. Application outcomes reshape cognitive state, reshaped cognitive state constrains execution, and execution outcomes refine the knowledge foundation, producing a continuously self-improving system whose improvement is governed at every stage.
Why Bidirectionality Is Required
The bidirectionality of the engine is required for the structural isomorphism between the platform's behavioral dynamics and human cognitive dynamics. Human cognition is characterized by circular causation: emotion influences judgment and judgment influences emotion; confidence influences action and the outcomes of action influence confidence; values constrain imagination and the products of imagination reshape the salience of values. A unidirectional architecture, whether a pipeline or a directed acyclic graph, cannot produce these circular dynamics. The engine's bidirectional pathways implement the same circular causation computationally.
No single pathway produces human-relatable behavior in isolation. The affect-to-confidence pathway alone produces dispositionally modulated hesitation but not the integrity-driven self-correction that follows deviation. The integrity-to-confidence pathway alone produces normatively constrained action but not the affectively modulated empathy that determines the moral cost of deviation. The confidence-to-forecasting pathway alone produces pause-then-deliberate dynamics but not the personality-modulated speculation that shapes the character of deliberation. It is the simultaneous operation of all pathways, the fully coupled feedback system operating on all cognitive domains concurrently, that produces behavioral dynamics structurally isomorphic to human cognitive dynamics. This is the basis of the platform-level claim that the isomorphism cannot be reduced to any single primitive or subset.
Disclosure Scope
The cross-primitive coherence engine, comprising the network of bidirectional feedback pathways through which each cognitive domain's state modulates every other domain, the named coupling pathways and their behavioral consequences, the three-phase coherence control loop of detection (empathy), recording (integrity), and restoration (self-esteem), the early, mid, and late-stage coping intercepts and the integrity-collapse response, the complete mutation lifecycle in which every cognitive domain participates at defined stages, and the three concurrent self-improving loops, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart) in the platform synthesis chapter. This article describes that disclosed mechanism. The scope extends to embodiments operating with fewer than all cognitive domains, in which the engine substitutes policy-defined default coupling inputs for absent domains and the confidence governor reduces authorization proportionally to the missing governance coverage, and to substrate-agnostic embodiments in which the engine operates on any substrate supporting persistent agent state with deterministic state transition functions, provided the coupling structure remains bidirectional and governance-integrated.
