Mechanism
A planning graph is introduced in Chapter 4 of the cognition disclosure as a first-class cognitive structure within the semantic agent architecture. It is a mutable, memory-referenced, directed semantic structure that represents one or more hypothetical future states of the agent, the agent's environment, or both. Each planning graph comprises a root node representing the agent's current verified state and a plurality of branches, where each branch represents a distinct hypothetical trajectory: a sequence of speculative mutations, delegation outcomes, environmental transitions, or intent resolutions that the agent is evaluating as possible futures. A planning graph is not an execution plan, a schedule, or a commitment. It is a pre-execution construct that exists in a structurally distinct computational domain from the agent's verified execution memory.
Each branch encodes a defined set of attributes: a speculative mutation sequence describing the hypothetical state transitions the branch represents; a projected outcome characterizing the expected terminal state if the branch were executed; an affective reinforcement tag encoding the emotional valence associated with the branch given the agent's current affective state and the projected outcome's alignment with the agent's intent; a trust slope projection encoding the hypothetical trust slope trajectory that would result from executing the branch; a policy compatibility flag indicating whether the branch's speculative mutations are admissible under the agent's current policy configuration; and a branch classification label. Because planning graphs are instantiated through defined interfaces, governed by policy constraints, modulated by affective state, constrained by the integrity field, and recorded in lineage only when promoted, the disclosure distinguishes them from systems that treat planning as a stateless function call, a prompt-engineering technique, or an external orchestration layer.
Structural Separation from Verified Memory
Planning graphs are maintained in structural separation from the agent's verified execution memory. The disclosure characterizes this separation not as a software convention, a namespace distinction, or an access control policy, but as an architectural invariant enforced at the substrate level. Verified execution memory, comprising the committed values of all agent fields, the lineage of all governance-validated mutations, and the accumulated results of executed operations, occupies a distinct computational domain from the planning graph structures. No mechanism exists by which a planning graph branch can directly modify verified execution memory without passing through the governance-validated promotion pathway.
The separation serves several stated purposes. It ensures that speculative reasoning cannot contaminate verified state: an agent that constructs a branch projecting a successful outcome does not thereby acquire the outcome as verified memory. It allows the agent to maintain multiple contradictory hypothetical futures simultaneously, for instance one branch projecting task success and another projecting task failure, without producing internal inconsistency, because neither branch has been promoted. The separation is also bidirectional. When the forecasting engine constructs a planning graph, it reads the agent's current verified state as the root node but does not establish a live reference, so subsequent verified state changes do not automatically propagate into existing planning graphs. This snapshot isolation makes planning graph evaluations deterministic with respect to the verified state at the time of graph construction.
The Forecasting Engine
The forecasting engine is a substrate module instantiated at the agent level or zone level that constructs, evaluates, modulates, and manages planning graphs throughout their lifecycle. It is not an external service, a shared utility, or a centralized scheduler. It is a component of the agent's own cognitive substrate that operates on the agent's own state, subject to the agent's own policy constraints, and modulated by the agent's own affective and integrity fields. The disclosure describes the engine in terms of five principal components.
Planning graph instantiation logic creates new planning graphs from the current verified state, reading the intent field to determine objectives, the context block to determine environmental conditions, and the memory field to identify historical patterns and prior planning graph outcomes that should inform branch construction. An affective prioritization module orders and weights branches based on the agent's current affective state, elevating branches whose projected outcomes align with the agent's disposition and deprioritizing those that conflict. A slope validation module evaluates each branch against the agent's trust slope trajectory, computing a hypothetical Derived Anchor Hash representing the trust slope state that would result if the branch were promoted. A personality-based modulation filter adjusts construction and evaluation parameters based on the agent's personality field. A pruning manager removes branches that are no longer viable, relevant, or computationally justified.
The Forecasting Execution Cycle
The forecasting engine operates through a defined execution cycle invoked at each cognitive decision point, when the agent must evaluate candidate actions, select among alternatives, or determine whether to act, delegate, or defer. The disclosure describes the cycle as synchronous within the deliberation pipeline rather than a background or periodic batch operation, and specifies six sequential phases.
In initialization, the engine reads current verified state and constructs or refreshes a planning graph rooted at that state. In speculative mutation simulation, the engine applies each branch's hypothetical mutations to a sandboxed copy of the agent's state, never to verified execution memory, and computes the projected outcome. The disclosure notes the simulation is deterministic: the same input state and mutation sequence produce the same projected outcome, which it distinguishes from statistical tree search methods such as Monte Carlo Tree Search that evaluate over random rollouts. In slope projection and validation, the slope validation module computes the hypothetical Derived Anchor Hash and confirms whether each branch maintains trust slope continuity. In the policy compatibility check, each slope-eligible branch is evaluated against the current policy configuration. In emotional reinforcement tagging, each slope-eligible, policy-compatible branch receives an affective reinforcement tag. In branch marking and pruning, each branch receives a classification label, and branches that fail slope validation or policy compatibility are marked for removal.
Slope-Constrained Simulation
The speculative simulation is slope-constrained: the trust slope trajectory serves as a structural filter determining which hypothetical futures the agent is permitted to evaluate for promotion. The disclosure frames this as a hard architectural boundary rather than a soft preference or ranking criterion. For each branch, the slope validation module applies the branch's speculative mutation sequence to a sandboxed copy of the agent's lineage and computes the trust slope hash that would result. That hypothetical Derived Anchor Hash is compared against the agent's current trust slope trajectory using the same continuity validation algorithm the governance infrastructure applies to committed mutations. If the hypothetical hash maintains continuity, the branch is slope-eligible; if it would produce a lineage gap, a hash chain discontinuity, or a provenance violation, the branch is slope-ineligible.
The constraint operates prospectively, filtering branches before they reach the promotion interface, so the governance pipeline never receives a promotion candidate that would fail trust slope validation. Only slope-eligible branches may be promoted to execution. A slope-ineligible branch may be retained for introspective purposes, enabling the agent to understand why certain hypothetical futures are structurally foreclosed, but it cannot advance through the promotion interface. The disclosure states that this ensures the forecasting engine, regardless of its speculative breadth, cannot produce execution candidates that would violate the system's trust and provenance guarantees.
Branch Classification
Each branch is assigned a classification label drawn from a four-category taxonomy that determines the branch's role and the operations that may be performed on it. An eligible branch has passed slope validation, satisfied policy compatibility, and received positive or neutral affective reinforcement; it is a viable candidate for promotion, ranked by a composite score comprising projected outcome quality, trust slope continuation magnitude, integrity impact projection, affective reinforcement strength, and alignment with current intent. An introspective branch has passed slope validation and policy compatibility but received negative affective reinforcement; it is retained for cognitive self-examination rather than promotion, enabling the agent to reason about why certain futures are aversive and to detect affective biases that may be distorting its planning.
A delegable branch is slope-eligible and policy-compatible but represents a trajectory the agent's policy configuration or personality field identifies as better suited for delegation to a child agent. A pruned branch has failed slope validation, failed policy compatibility, exceeded the pruning manager's thresholds, or been superseded by a higher-ranked branch. The disclosure states that classification is not permanent: an introspective branch may become eligible if the agent's affective state shifts, an eligible branch may become pruned if supporting environmental conditions change, and a delegable branch may become eligible if the delegation target is unavailable. The forecasting execution cycle re-evaluates classifications at each iteration so the planning graph reflects the current cognitive landscape.
The Containment Layer and Delusion Boundary
The containment layer is the structural enforcement mechanism that maintains the separation between the speculative planning graph domain and verified execution memory. The disclosure describes it as an architectural boundary embedded in the cognitive substrate, not a software flag, metadata annotation, or runtime check, that prevents speculative content from being treated as verified reality except through governance-validated promotion. It enforces several invariants simultaneously: every data element in a planning graph carries an immutable speculative marker at construction that only the promotion interface may strip; read isolation prevents the agent's verified execution processes from accessing planning graph content as if it were verified memory; and speculative content is not written to lineage as committed state, since lineage records only governance-validated mutations.
The disclosure defines a delusion boundary condition: a formally specified pathological state in which the containment layer fails and speculative content is treated as verified reality. It names this containment collapse and calls it the architectural analog of delusion, the condition in which the agent can no longer distinguish what it has speculatively projected from what has actually occurred. Failure modes include corruption or stripping of the speculative marker without promotion, a breach of read isolation, and admission of speculative content through the promotion interface without completing governance validation. The system provides containment integrity verification mechanisms, including periodic containment audits, boundary crossing monitors, lineage consistency checks, and behavioral coherence monitors, and on detection initiates a containment restoration protocol that suspends execution authority, quarantines affected structures, performs lineage forensic analysis, reconstructs verified state from the most recent governance-validated checkpoint, and re-initializes the containment layer.
Personality and Affective Modulation
The personality field is a structured data object comprising trait dimensions that deterministically shape the forecasting engine's instantiation logic, branch generation parameters, and evaluation criteria. The disclosure enumerates trait dimensions including risk tolerance, introspective depth, impulsivity, fallback rigidity, delegation preference, and temporal planning horizon, each modulating how the engine generates and evaluates branches. The personality field may be configured statically, adapted within policy-defined bounds based on accumulated outcomes, or evolved over time through a feedback mechanism, with the configuration mechanism specified by the policy reference field and the evolution history recorded in lineage.
The affective state field separately modulates planning graph construction through defined coupling pathways, encoding the agent's current, rapidly-changing orientation rather than the personality field's slowly-evolving disposition. Current risk sensitivity and novelty appetite determine planning graph expansion depth, with elevated risk sensitivity producing shallower, higher-confidence branches and elevated novelty appetite producing deeper, more exploratory branches. Affective state biases branch prioritization, influences the rate at which branches are classified as delegable, and governs how long partially-failed branches are retained before pruning. The disclosure is explicit that this modulation respects the governance separation: affective state shapes how planning graphs are constructed and evaluated but does not determine whether branches are admissible for promotion, which remains identical regardless of affective state.
Executive Graph Aggregation
The executive engine is a substrate module that aggregates planning graphs from a plurality of agents within a shared operational scope, such as a zone or delegation hierarchy, into a unified executive graph representing the collective speculative state of the multi-agent system. The disclosure distinguishes two structural tiers: micro-planning graphs, the agent-level structures described above, and macro executive graphs, zone-level or group-level structures synthesized by aggregating, aligning, and reconciling the micro-planning graphs. The executive graph is not a simple union; it identifies branch intersections, pairs or groups of branches from different agents that reference the same resources, target the same delegation endpoints, or project outcomes that depend on other agents, which are the structural basis for coordination.
The executive graph arbitrates among agents' planning graphs using three criteria applied in priority order: slope compatibility first, emotional reinforcement alignment second, and personality profile alignment third. When standard arbitration produces an inconclusive result, the disclosure describes an emotional quorum override, a tiebreaker in which a policy-defined supermajority of affected agents exhibiting strong positive affective reinforcement toward a branch resolves the conflict, operating within governance constraints rather than overriding them. The executive graph maintains its own containment layer, structurally separate from those of the individual agents, so zone-level speculative coordination does not contaminate zone-level verified state. The disclosure frames this architecture as a replacement for centralized scheduling: an agent begins executing a task not because a scheduler assigned it, but because the agent's own forecasting engine generated a branch, classified it as eligible, and promoted it through the governance-validated promotion interface.
Disclosure Scope
The disclosure covers the planning graph as a first-class cognitive structure, mutable, memory-referenced, and directed, rooted at the agent's current verified state with branches encoding speculative mutation sequences, projected outcomes, affective reinforcement tags, trust slope projections, policy compatibility flags, and classification labels. It covers the structural separation of planning graphs from verified execution memory with bidirectional snapshot isolation; the forecasting engine and its five components; the six-phase forecasting execution cycle; slope-constrained speculative simulation through the hypothetical Derived Anchor Hash; the eligible, introspective, delegable, and pruned branch taxonomy; the containment layer and the delusion boundary condition with its restoration protocol; personality-based and affective modulation under governance separation; and executive graph aggregation, arbitration, and the emotional quorum override. These mechanisms are disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart), in Chapter 4. This article describes that disclosed mechanism. The scope extends to centralized, federated, decentralized, and embodied deployments in which the governance requirements, promotion interface semantics, and containment layer enforcement remain invariant.
