Non-Executing Cognitive Mode

Mechanism

The Non-Executing Mode (NEM) mechanism is defined in Chapter 5 of the cognition patent as a first-class operating state of the agent runtime, on equal structural footing with the executing state rather than as an exception path or interrupt handler. Entry to NEM is triggered by a confidence collapse predicate evaluated continuously over the agent's confidence vector. When the predicate fires — for example, when a calibrated confidence score falls below a hysteretic floor, or when a forecast disagreement metric exceeds a configured ceiling, or when a contradicting observation invalidates a load-bearing precondition — the runtime transitions atomically into NEM and emits a durable lineage record describing the trigger, the state at trigger time, and the policy reference that governed the transition.

Once in NEM, the agent's cognitive faculties remain fully active. Perception continues to ingest observations. The forecasting engine continues to produce branches, classify them, and update its planning graph. The inquiry subsystem can still issue questions, request clarification, and consult human or peer authorities. What is structurally suspended is the agent's ability to commit to the world: every effector, tool invocation, write to verified memory, outbound communication channel, and external API binding is gated behind an executing-state guard. The guard is not a soft check inside business logic; it is a precondition asserted at the dispatch boundary of the action runtime, and any attempt to bypass it produces a recorded violation rather than a side effect.

NEM is durable. Unlike an interrupt or transient pause, the state persists across the agent's normal cognitive cycles, across restarts of the runtime, and across handoffs between operating contexts. The state is materialized in the same canonical fields that record commitments and policies, which means that any downstream observer — auditor, supervising agent, or operator — can determine the agent's executing status without privileged access to internal flags. Re-entry to the executing state is not automatic and is not a function of elapsed time. It requires an explicit re-entry path: a policy-defined sequence of conditions that must all be satisfied, including renewed confidence above an exit threshold with margin, resolution of the originating trigger, and, where the policy requires it, an external acknowledgment from a designated authority. The re-entry path is itself recorded in lineage, so that the conditions under which the agent resumed acting are reconstructable after the fact.

Operating Parameters

The trigger predicate operates over inputs drawn from the confidence governance subsystem, including a scalar confidence value computed deterministically from current canonical fields, first and second derivatives of that value to detect adverse trajectories before the floor is crossed, and a set of named invalidation events that can force entry independent of the scalar (for instance, a contradiction signal from the forecasting engine or a hard precondition violation reported by a tool). Each of these inputs is configurable per deployment, and the predicate itself is expressed in the agent's policy reference rather than compiled into the runtime, so domains with different risk profiles can adopt the same architectural mechanism without forking the implementation.

Hysteresis margins separate the entry threshold from the exit threshold to prevent oscillation near the boundary. The entry margin is conservative and biases the agent toward NEM when ambiguity is present; the exit margin is stricter and requires sustained confidence above a higher level for a configured dwell time before re-entry to the executing state is permitted. The dwell time is itself a parameter, ranging from sub-second values appropriate for low-stakes interactive tasks to multi-minute or multi-hour values appropriate for clinical, financial, or safety-critical contexts.

Within NEM, the policy reference may further restrict cognitive activity. A minimal-NEM configuration permits all internal cognition while blocking only effectors. A reflective-NEM configuration additionally restricts the forecasting engine to corroborating and contradictory branches, suppressing speculative extrapolation until trust is restored. A custodial-NEM configuration narrows the inquiry channel to a single designated authority. These configurations are selected by policy and recorded with the state, so that the operating envelope of the agent during NEM is fully specified rather than implicit.

Re-entry preconditions are expressed as a conjunction of clauses: confidence-restored, trigger-resolved, dwell-elapsed, and, optionally, authority-acknowledged. Each clause evaluates against canonical fields and produces a boolean result with a justification record. The conjunction must hold before the runtime issues an executing-state transition, and the justification record for each clause is preserved so that re-entry, like entry, is auditable.

Alternative Embodiments

In a single-agent deployment, NEM is realized as a state flag on the runtime that gates the action dispatcher. In a multi-agent deployment, NEM may be propagated to peer agents that depend on the focal agent's commitments, allowing them to enter their own protective states or substitute alternate authorities. In a hierarchical deployment, an outer supervising agent may be the designated authority for re-entry acknowledgment, producing a chain of NEM custody where each layer's resumption is contingent on the layer above.

Alternative trigger embodiments include statistical change-point detectors over observation streams, model-disagreement metrics computed across an ensemble of forecasters, and externally injected NEM commands from a human operator or regulator. Alternative re-entry embodiments include time-locked re-entry where dwell-elapsed is the dominant clause, evidence-locked re-entry where a specific corroborating observation must arrive, and quorum re-entry where multiple authorities must independently acknowledge before resumption.

The mechanism is also embodied in environments that lack a single privileged runtime, such as distributed agent meshes. In such environments, NEM is realized as a distributed lock over the effector namespace, with the same structural guarantees: cognition continues, effectors are blocked, the state is durable, and re-entry is governed by an explicit policy-defined path.

Composition With Other Mechanisms

NEM composes directly with the confidence governance pipeline that produces its trigger inputs and with the forecasting engine whose branch classifications can drive contradiction-based entries. It composes with the lineage subsystem, which records every entry, exit, and intra-NEM cognitive event in canonical form, and with the policy reference subsystem, which supplies the configurable predicates and re-entry conditions. Because NEM is a structural state rather than a heuristic, it composes additively with any other mechanism that already respects the executing-state guard at its dispatch boundary. The composition is monotonic: adding NEM to a system that already satisfies its dispatch invariants does not reduce that system's expressive power within the executing state, and removing NEM from such a system does not introduce defects within the executing state — it only removes the protective regime that NEM provides.

In particular, NEM composes with the branch classification system: while in NEM, contradictory and corroborating branches continue to be evaluated and may themselves resolve the originating trigger, providing a path back to the executing state through evidence rather than time. NEM also composes with the memory separation mechanism: forecast memory updates remain admissible during NEM, but no promotion to verified memory or action memory may occur, preserving the structural isolation that prevents speculative contamination during a low-confidence interval.

Distinction From Prior Art

Prior systems addressing low-confidence operation generally adopt one of three strategies: full halt, full proceed with logged warnings, or human-in-the-loop confirmation prompts. Full halt sacrifices situational continuity and forfeits the cognitive work the agent could still safely perform. Full-proceed-with-warnings is unsafe in any domain where the warning consumer is itself the agent. Human confirmation introduces latency that is unacceptable in many real-time contexts and does not scale to autonomous operation.

NEM is structurally distinct from each of these. It is not a halt because cognition continues and the agent remains a useful observer and forecaster of the environment. It is not a proceed because effectors are architecturally gated rather than warned about. It is not a confirmation prompt because the re-entry path is policy-defined and may include autonomous evidence-based clauses that do not require synchronous human input. Prior reinforcement-learning safety techniques such as constrained optimization or shielded policies operate within the action policy itself; NEM operates outside it, at the dispatcher, and so cannot be circumvented by a policy that has learned to evade its own constraints.

Disclosure Scope

The disclosure encompasses any agent runtime that maintains a durable non-executing state with the following structural properties: full cognitive operation including forecasting and inquiry; architectural prohibition of effector commitment at the dispatch boundary; policy-defined entry predicate over confidence and contradiction signals; policy-defined re-entry path with at least one non-temporal clause; and lineage recording of entry, intra-state events, and exit. The disclosure is not limited to any particular embodiment of confidence scoring, predicate language, or runtime architecture, and it covers single-agent, multi-agent, hierarchical, and distributed deployments.

The disclosure further extends to embodiments in which the dispatch boundary is realized in software at the runtime layer, in middleware between the agent and its environment, in sandbox supervisors, in orchestration controllers, or in hardware protection domains. In each case the structural property is the same: the prohibition on effector commitment is enforced at a layer the agent's decision policy cannot reach, and is therefore robust against learned evasion. The disclosure also covers embodiments in which NEM is composed with other supervisory states — for example, dormant, observational, or shadow modes — provided that the durable, audited, policy-governed entry and re-entry properties are preserved.

Because the mechanism is policy-governed and deterministic, it can be formally analyzed, audited, and certified. Regulatory compliance is demonstrable through structural analysis of the dispatch boundary and the re-entry conjunction rather than solely through empirical testing. The conjunction is in particular a tractable verification target: each clause refers to canonical fields whose values at any moment are recoverable from lineage, so the question of whether a particular re-entry was justified reduces to the question of whether each clause's record corresponds to its policy specification at the moment of evaluation. Different domains tune the parameters through policy configuration without architectural change, making the same disclosed mechanism applicable to autonomous vehicles, companion AI, therapeutic agents, and enterprise systems. The disclosure additionally encompasses methods of verifying NEM compliance through replay of recorded lineage against policy specifications, methods of certifying NEM behavior under adversarial trigger sequences, and methods of composing NEM with regulatory escalation paths in jurisdictions that require human review of autonomous suspension events.