Task Class Differentiation Under Confidence Interruption

Mechanism

Task class differentiation is the part of confidence-governed execution that adapts the response to a confidence interruption based on the structural character of the task the agent was running when its confidence collapsed. The premise of the surrounding architecture is that execution is a revocable permission, not a default state: a confidence governor continuously computes the agent's assessed sufficiency and, when that sufficiency falls below the execution authorization threshold or when the confidence trajectory triggers a preemptive suspension, withdraws authorization and moves the agent into a suspended state in which cognition continues but action does not. The question task class differentiation answers is what the agent should do at that moment, given that not all tasks are equivalent in their reversibility, their tolerance for partial execution, or their amenability to alternative continuation strategies.

Rather than apply a single uniform halt to every interrupted task, the confidence governor recognizes at least three structurally distinct task classes and applies a differentiated interruption protocol to each. The class is not a label chosen by the agent's intent; it is assigned by a task class classifier that evaluates the task's structural properties, and the assignment is recorded in the agent's lineage so that the interruption decision is auditable by governance infrastructure.

The Terminal Task Class

The terminal task class comprises tasks characterized by high irreversibility, high cost of partial execution, and low tolerance for state corruption. The spec gives operations that commit permanent changes to external systems, transactions that cannot be rolled back, physical actions with irreversible consequences, and communications that cannot be retracted once transmitted as examples of this class.

When the confidence governor suspends execution during a terminal task, the interruption protocol prioritizes state preservation and partial progress protection. The agent preserves the current execution state, including all uncommitted intermediate results, all acquired locks or reservations, and all accumulated context, in a durable, governance-tagged checkpoint that can be restored when execution authorization is recovered. The agent does not attempt to redirect, explore alternatives, or perform creative reinterpretation of the task. It halts execution at the earliest safe point and protects the partial progress achieved so far. The disposition for a terminal task is conservative by construction: the worst outcome to avoid is a half-committed irreversible action.

The Exploratory Task Class

The exploratory task class comprises tasks characterized by low irreversibility, low cost of partial execution, and high tolerance for redirection. The spec names search operations, information gathering, hypothesis testing, comparative analysis, and open-ended investigation as members of this class.

When the confidence governor suspends execution during an exploratory task, the interruption protocol does not freeze the task in place. Instead it redirects the agent's cognitive capacity toward hypothesis expansion rather than state preservation. The agent broadens its search space, generates alternative hypotheses, explores adjacent problem formulations, and evaluates previously unconsidered approaches. Because the partial work of an exploratory task is cheap to discard and nothing irreversible is at stake, the suspension becomes an opportunity to widen the inquiry rather than a loss to be protected.

The Generative Task Class

The generative task class comprises tasks characterized by creative or inventive objectives, moderate irreversibility, and high sensitivity to commitment timing. The spec names content creation, design synthesis, solution invention, and any task in which the agent produces novel output rather than executing a defined procedure.

When the confidence governor suspends execution during a generative task, the interruption protocol transitions the agent to a lower-commitment creative exploration mode. The agent shifts from producing finished output to generating prototypes, sketches, partial formulations, and tentative hypotheses. It does not commit to any single creative direction. It generates a plurality of candidate directions and evaluates them comparatively without finalizing any. The sensitivity of a generative task is to premature commitment, so the protocol holds the agent in an uncommitted, optionality-preserving posture until authorization returns.

The Task Class Classifier

Task class assignment is determined by a task class classifier that evaluates the task's structural properties, including irreversibility magnitude, partial execution cost, redirection tolerance, and commitment sensitivity, and assigns the task to one of the three classes. The classifier may also assign a task to a hybrid class combining elements of two or more base classes. When it does, the interruption protocol inherits the most conservative constraints from each contributing class, so a task that is partly terminal and partly exploratory is treated with the caution its terminal aspect demands. Task class assignment is recorded in the agent's lineage and is auditable by governance infrastructure, which means the basis on which an interrupted task was handled can be reconstructed after the fact rather than inferred from behavior.

Composition With the Surrounding Governor

Task class differentiation does not operate in isolation. It is invoked by the confidence governor at the moment authorization is withdrawn, and the withdrawal itself follows from the governor's continuous confidence computation, which draws on capability sufficiency, resource availability, internal integrity state, affective modulation, and the task and environment state. The differentiated protocols also feed the recovery path described in the same chapter: when execution authorization is later restored through confidence restoration, stability verification, and reauthorization, the agent incorporates the products of suspension-time cognition into its resumed plan. A terminal task resumes from its governance-tagged checkpoint, an exploratory task resumes with a broadened hypothesis set, and a generative task resumes with the plurality of candidate creative directions it generated while suspended.

The classes also condition the effort analysis the chapter describes. For terminal tasks the effort analysis weights reliability and safety more heavily than effort minimization, preferring a higher-effort path with lower risk over a lower-effort path with higher risk. For exploratory tasks it weights breadth of exploration, preferring paths that cover more of the search space. For generative tasks it weights creative optionality, preferring paths that preserve more degrees of freedom. The same class assignment that selects the interruption protocol therefore also shapes how the agent evaluates its way back to execution.

Distinction From Undifferentiated Interruption

Conventional autonomous agent systems, including runtime environments that provide pause and resume capabilities, suspend execution reactively in response to external failures or resource interruptions, and they suspend uniformly. An interrupted task is simply stopped. The mechanism here suspends proactively, on the agent's own continuously computed assessment of its sufficiency, and it does not treat every interrupted task identically. By recognizing that an irreversible commitment, an open-ended search, and a creative synthesis each require a different response to the same drop in confidence, the governor avoids the two failure modes of undifferentiated handling: freezing an exploratory task that would have benefited from widening, or redirecting a terminal task whose half-completed state should have been protected. The differentiation is what lets a single agent host a heterogeneous mix of work and respond to a confidence interruption without collapsing into either halt-everything or continue-everything behavior.

Disclosure Scope

Task class differentiation under confidence interruption, comprising the terminal, exploratory, and generative task classes, the task class classifier that assigns a task to one of those classes or to a hybrid class inheriting the most conservative constraints of its contributors, the differentiated interruption protocols of state preservation for terminal tasks, hypothesis expansion for exploratory tasks, and lower-commitment creative exploration for generative tasks, and the recording of the class assignment in the agent's lineage for governance audit, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart) in the confidence-governed execution chapter. This article describes that disclosed mechanism. The scope extends to embodiments in which the differentiated protocols feed the recovery and effort analysis paths of the same governor, and is independent of the specific task domain and of the specific substrate on which the agent executes.