Mechanism

Branch classification is the labeling step by which the forecasting engine assigns each branch of a planning graph a role in the agent's cognitive process and the set of operations that may be performed on it. A planning graph is a mutable, memory-referenced, directed structure whose root node represents the agent's current verified state and whose branches each represent a distinct hypothetical trajectory: a sequence of speculative mutations, delegation outcomes, environmental transitions, or intent resolutions that the agent is evaluating as a possible future. Classification is the final phase of the forecasting execution cycle, performed after a branch has been simulated, slope-validated, policy-checked, and emotionally tagged, and it determines whether the branch is a candidate for promotion, an object of self-examination, a unit of work to hand to a child agent, or material to remove.

The classification taxonomy comprises four categories: eligible, introspective, delegable, and pruned. The label a branch receives is a function of three prior evaluation results carried on the branch itself: whether it is slope-eligible, whether it is policy-compatible, and the sign of its affective reinforcement tag. The label is not an independent judgment layered on top of the branch; it is the structural summary of the evaluations the branch has already passed or failed, which is why the same four labels suffice across every consumer of the planning graph.

The Four Categories

A branch is classified eligible when it has passed slope validation, satisfied policy compatibility requirements, and received positive or neutral affective reinforcement. An eligible branch is a viable candidate for promotion to verified execution: it represents a hypothetical future the agent may choose to realize. Eligible branches are ranked by a composite score comprising the branch's projected outcome quality, its trust slope continuation magnitude, its integrity impact projection, its affective reinforcement strength, and its alignment with the agent's current intent. The highest-ranked eligible branch at a given evaluation point is the agent's leading candidate for execution, subject to the promotion interface's governance evaluation.

A branch is classified introspective when it has passed slope validation and policy compatibility but received negative affective reinforcement: its projected outcome is emotionally aversive to the agent given its current affective state. Introspective branches are not candidates for promotion. They are retained for cognitive self-examination, enabling the agent to reason about why certain futures are aversive, to detect affective biases that may be distorting its planning, and to surface branches that may become eligible if the agent's affective state changes. Retaining negatively reinforced but structurally viable branches is the mechanism by which the forecasting engine supports self-reflective cognition.

A branch is classified delegable when it is slope-eligible and policy-compatible but represents a trajectory better suited for transfer to a child agent. A branch may be delegable because it requires capabilities the agent does not possess but that are available through delegation, because it involves a sub-problem within the specialization domain of an available delegate, or because the agent's personality field, specifically the delegation preference trait, indicates the agent should delegate rather than execute the branch directly. Delegable branches are transferred to the planning graph delegation mechanism.

A branch is classified pruned when it has failed slope validation, failed policy compatibility, exceeded the pruning manager's entropy or compute thresholds, or been superseded by a higher-ranked branch that renders it redundant. Pruned branches are scheduled for removal by the pruning manager. Before removal they are briefly retained with their rejection annotation, so the agent can reference the reason for pruning in subsequent planning cycles. Once the retention period expires, the pruned branch is deleted, freeing computational resources.

How Labels Are Assigned

Classification is the sixth and final phase of the forecasting execution cycle, which is invoked synchronously at each cognitive decision point. The preceding phases establish the inputs that classification reads. In the speculative mutation simulation phase, the branch's mutations are applied to a sandboxed copy of the agent's state, never to verified execution memory, and a projected outcome is computed deterministically. In the slope projection and validation phase, the slope validation module computes a hypothetical Derived Anchor Hash (DAH') for the branch and compares it against the agent's current trust slope trajectory: branches that would produce a trust slope discontinuity are flagged slope-ineligible, and branches that maintain continuity are confirmed slope-eligible. In the policy compatibility phase, each slope-eligible branch is checked against the agent's current policy configuration, with branches containing policy-excluded mutations flagged policy-incompatible. In the emotional reinforcement tagging phase, each surviving branch receives an affective reinforcement tag encoding how well its projected outcome aligns with the agent's current affective disposition.

The marking phase then reads these results and assigns the label. Branches that are slope-eligible, policy-compatible, and positively reinforced are marked eligible. Branches that are slope-eligible and policy-compatible but negatively reinforced are marked introspective. Branches that are slope-eligible, policy-compatible, and suitable for transfer to a child agent are marked delegable. Branches that fail slope validation, policy compatibility, or both are marked pruned and scheduled for removal. Because the label is derived from these structural evaluations rather than from the branch's surface content, the same classification logic applies uniformly across decision contexts.

Classification Is Not Permanent

A branch's classification reflects the agent's state at the moment of evaluation, and the forecasting execution cycle re-evaluates branch classifications at each iteration so the planning graph continues to reflect the agent's current cognitive landscape. An introspective branch may become eligible if the agent's affective state shifts so that the branch's projected outcome is no longer aversive. An eligible branch may become pruned if the environmental conditions that supported it change. A delegable branch may become eligible if the delegation target becomes unavailable and the agent must execute the branch itself.

This dynamism is the basis for branch reinterpretation, in which a dormant or active branch is re-evaluated under changed conditions and assigned a new classification or projected outcome. A branch that was introspective because it was negatively reinforced may be reinterpreted as eligible when the agent's affective state shifts; a branch that was slope-ineligible may be reinterpreted as slope-eligible when the trust slope trajectory changes; a branch that was policy-incompatible may be reinterpreted as policy-compatible when the policy configuration is updated. Reinterpretation, together with dormancy and deferred promotion, lets the forecasting engine manage branches across temporal horizons that exceed any single evaluation cycle.

Classification Within the Speculative Domain

Every branch a classification label is assigned to lives in the planning graph domain, which is held in structural separation from the agent's verified execution memory. This separation is an architectural invariant enforced at the substrate level, not a software convention or an access control policy. No mechanism exists by which a branch, regardless of its classification or evaluation score, can directly modify verified execution memory. The sole pathway from speculative to verified status is the promotion interface, a governance-controlled gateway that subjects a candidate branch to the full governance evaluation pipeline and either admits it as a committed mutation or returns it to the speculative domain with a rejection annotation.

The containment layer that enforces this separation tags every element of a planning graph, including each classification label, with an immutable speculative marker at construction time. Only the promotion interface, upon successful governance validation, strips that marker and re-tags content as verified. Branches do not produce lineage entries while they remain speculative; the lineage records governance-validated mutations, though it may record metadata about the forecasting process itself, such as the creation, evaluation, and pruning of branches as cognitive events. An eligible label is therefore never a substitute for promotion: it identifies a candidate, and the candidate still has to pass the promotion interface before any of its content reaches verified memory.

What Shapes the Distribution of Labels

Two structural inputs shape how branches are generated and evaluated, and therefore how labels are distributed across a planning graph, without changing the criteria that define each label. The personality field is a structured object of trait dimensions encoding the agent's characteristic, slowly evolving approach to speculative reasoning. Its delegation preference trait widens or narrows the set of branches classified delegable rather than eligible; its risk tolerance, impulsivity, fallback rigidity, introspective depth, and temporal planning horizon traits adjust how branches are generated, how long they are retained, and how deeply they are evaluated before classification.

The affective state field encodes the agent's current, rapidly changing disposition based on recent outcomes and observations, and it biases branch prioritization and retention. Branches whose projected outcomes align with the agent's current affective disposition receive higher priority in the evaluation queue, and the agent's persistence-under-partial-failure value determines how long partially failed branches are retained before being reclassified as pruned. Affective modulation shapes how the forecasting engine constructs and evaluates branches but does not determine admissibility for promotion: the governance requirements for promotion remain identical regardless of the agent's affective state. Personality and affect move where the labels land; they do not redefine what the labels mean.

Distinction From Prior Art

Conventional planning systems treat forward simulations as homogeneous nodes in a search tree, distinguished by score but not by role. Statistical tree search methods such as Monte Carlo Tree Search evaluate nodes over random rollouts rather than producing a structural label that ties each branch to its slope eligibility, policy compatibility, and affective reinforcement. The branch classification taxonomy instead assigns each branch one of four roles derived deterministically from these prior evaluations, and it keeps negatively reinforced but viable branches as a first-class introspective category rather than discarding them, supporting self-examination that score-only search trees do not represent.

Because classification governs which operations a branch admits, and because every classified branch remains in the speculative domain behind the promotion interface, the taxonomy is the structural variable that separates the kinds of reasoning that may inform commitment from the kinds that may not. An eligible label marks a candidate for governed promotion; an introspective, delegable, or pruned label keeps a branch out of the promotion path while preserving it for reflection, delegation, or removal.

Disclosure Scope

The branch classification taxonomy, comprising the eligible, introspective, delegable, and pruned categories, the assignment of each label from a branch's slope eligibility, policy compatibility, and affective reinforcement during the marking phase of the forecasting execution cycle, the re-evaluation and reinterpretation of classifications as the agent's state evolves, and the maintenance of every classified branch within the speculative planning graph domain behind the governance-validated promotion interface, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart) in Chapter 4 at Section 4.6, with the surrounding execution cycle, slope constraint, containment layer, and personality and affective modulation disclosed in Sections 4.4, 4.5, 4.7, 4.8, and 4.9. This article describes that disclosed mechanism. The scope extends to embodiments in which classification is re-evaluated at each forecasting iteration and in which the distribution of labels is modulated by the personality and affective state fields, provided the four categories and their derivation from slope, policy, and reinforcement evaluation are preserved.