Mechanism
Hysteretic recovery is the protocol by which a semantic agent regains execution authorization after the confidence governor has suspended it. As disclosed in Chapter 5, execution is a revocable permission rather than a default state: the confidence governor continuously computes a confidence value encoding the agent's assessed sufficiency to continue its task, and it gates execution against a single authorization threshold. When confidence falls below that threshold, the agent transitions from the authorized state to the suspended state and is structurally prohibited from committing mutations. Hysteretic recovery governs the reverse transition, from suspended back to authorized, so that the return to execution is not the mirror image of the departure from it.
The defining property is asymmetry at the threshold. The transition from authorized to suspended occurs when confidence crosses below the authorization threshold. The transition from suspended back to authorized does not occur at the same value. Recovery requires that the confidence value exceed the authorization threshold by a configurable hysteresis margin, and that it remain above that margin throughout a stability verification period before reauthorization is granted. Confidence that merely returns to the threshold, or that crosses it briefly, does not restore execution. This asymmetry is the hysteresis: the path back to execution sits above the path out of it.
The purpose of the asymmetry is to prevent oscillation. A gate that suspended and reauthorized at a single shared value would flip on every cycle whenever confidence fluctuated near that value, leaving the agent toggling between executing and non-executing modes. The hysteresis margin removes that failure mode by construction: once suspended, the agent must achieve and hold a meaningfully higher confidence before it executes again, so noise that does not clear the margin cannot trigger a transition.
The Three-Phase Recovery Process
Recovery of execution authorization proceeds through three phases in sequence: confidence restoration, stability verification, and reauthorization. The agent does not resume execution prematurely or under conditions that would immediately re-trigger suspension, because each phase must complete before the next begins.
Confidence restoration is the phase in which the agent's confidence value increases from below the authorization threshold back above it. Restoration may result from resolution of the adverse conditions that caused the original drop, from resource restoration, capability acquisition, uncertainty reduction, or integrity repair, from successful completion of inquiry operations that supplied information enabling higher confidence, from changes in task state that reduced the task's demands relative to the agent's capabilities, or from changes in environmental conditions that mitigated adverse factors. Restoration is computed by the same confidence evaluation function used during normal operation, applied to the agent's updated state. It is not a separate or relaxed computation.
Stability verification follows restoration and precedes reauthorization. During this phase the confidence governor monitors both the confidence value and the confidence trajectory across a configurable verification period to confirm that the restored confidence is stable: that the value is not fluctuating near the threshold, that the differential rate is not trending negatively, and that no alarm conditions are active. This phase exists to catch transient restoration, where a temporary improvement in conditions produces a brief confidence spike that quickly decays. A spike that does not persist across the verification window does not reach reauthorization.
Reauthorization is the act by which the confidence governor restores the agent's execution pathway. Upon successful completion of stability verification, the governor transitions the agent from suspended to authorized, reconnects the execution subsystem's output pathway, and notifies the deliberation pipeline that execution is available. The agent then evaluates its current planning graph, selects the highest-ranked eligible branch, and resumes execution.
The Hysteresis Margin
The hysteresis margin is the configurable margin by which the confidence value must exceed the authorization threshold throughout the stability verification period. Its function is to ensure that the agent's confidence is not merely above the threshold but meaningfully above it, providing a buffer against immediate re-suspension. Reauthorization on a value that sits exactly at the threshold would leave the agent one fluctuation away from suspension again; the margin reserves headroom so that ordinary confidence variation after recovery does not push the agent back below the gate.
The magnitude of the margin is not fixed. It is configurable based on the task class, the severity of the original suspension event, and the duration of the suspension. Longer suspensions require larger hysteresis margins, because a longer suspension indicates more severe or persistent adverse conditions and therefore demands stronger evidence of recovery before execution resumes. The margin is thus responsive to the circumstances of the suspension it is recovering from, rather than a single constant applied uniformly.
Trajectory and Stability
Recovery is evaluated against the confidence trajectory, not only the instantaneous confidence value. The confidence evaluation function produces both a confidence value and a confidence rate of change, and the confidence governor performs differential rate analysis by computing the difference between the decay rate and the recovery rate at each evaluation cycle. A positive differential rate indicates that recovery is outpacing decay and the trajectory is upward; a negative differential rate indicates the opposite.
During stability verification, the governor uses this trajectory information to distinguish durable recovery from apparent recovery. Confidence that has crossed above the margin but whose differential rate is trending negative is recovering in value while deteriorating in direction, and the verification phase will not pass it. By requiring that the value hold above the margin and that the trajectory not be turning back toward suspension across the verification window, the protocol guards against reauthorizing at the apex of an unstable rise. This is the same trajectory awareness that, on the suspension side, lets the governor act before confidence crosses the threshold; on the recovery side it ensures the governor does not act until confidence is genuinely settled above it.
Composition with Suspension-Time Cognition
Hysteretic recovery composes with the structural separation between execution and cognition. Suspension withdraws the agent's ability to act but not its ability to think: in the non-executing cognitive mode, the agent continues forecasting, planning graph construction, inquiry generation, and delegation evaluation. The products of that suspension-time cognition feed directly into resumed execution. If task class differentiation produced a checkpointed state for a terminal task, a broadened hypothesis set for an exploratory task, or a plurality of candidate creative directions for a generative task during the suspension period, the agent incorporates these products into its resumed execution plan at reauthorization.
Recovery also composes with the confidence-integrity feedback loop. Because a suspended agent cannot commit mutations, it cannot commit integrity-violating mutations, so suspension creates a structural shield against further integrity degradation while restorative repair proceeds. The loop converges toward a state in which restored integrity supports restored confidence, which the recovery protocol then verifies for stability before reauthorizing. Where integrity is so severely degraded that no achievable confidence value can support reauthorization, a circuit breaker transitions the agent to the locked state, whose recovery requires external authorization rather than the autonomous hysteretic protocol.
Every phase of recovery is recorded in the agent's lineage alongside the confidence trajectory that produced it. Governance infrastructure can audit the recorded values to verify that reauthorization occurred only after the confidence value exceeded the threshold by the hysteresis margin throughout the verification period, and that no execution occurred while confidence was below the authorization threshold.
Prior-Art Distinctions
Selective prediction and abstention systems implement a binary abstain-or-predict decision based on a single confidence score from a model's own output distribution, and they resume prediction whenever that score recovers. The disclosed mechanism differs in that confidence is computed from multiple coupled cognitive domain fields, that the agent continues speculative reasoning during suspension rather than going idle, and that resumption is gated by a hysteresis-margin recovery protocol with a stability verification period rather than by a single recovered score.
Runtime environments that provide pause and resume capabilities suspend reactively in response to external failures or resource interruptions and resume when the external condition clears. The disclosed mechanism suspends proactively on the agent's own continuously computed sufficiency and recovers only after the agent verifies that its restored confidence is stable above the margin. Recovery is therefore governed by the agent's internal trajectory rather than by the mere removal of an external fault.
Disclosure Scope
Hysteretic recovery of execution authorization, comprising the asymmetric transition in which suspension occurs at the authorization threshold while reauthorization requires the confidence value to exceed that threshold by a configurable hysteresis margin held throughout a stability verification period, the three-phase recovery process of confidence restoration, stability verification, and reauthorization, the dependence of the margin's magnitude on task class and on the severity and duration of the suspension, the use of confidence trajectory and differential rate to distinguish durable from transient recovery, and the temporal reauthorization rule under which elapsed time alone does not restore authorization, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart) in Chapter 5. This article describes that disclosed mechanism. The scope is not limited to any particular confidence computation, any particular verification period, or any particular margin value, provided that reauthorization remains gated by a hysteresis margin verified for stability rather than by a single threshold crossing.
