Planning Graphs as First-Class Cognitive Structures

Mechanism

Planning graphs as first-class cognitive structures are defined in Chapter 4 of the cognition patent. A planning graph is a directed acyclic structure whose nodes denote candidate semantic states and whose edges denote candidate transitions, each annotated with the policy reference under which the candidate was generated. The forecasting engine maintains planning graphs in a memory region that is structurally distinct from verified execution memory; the boundary is enforced by the type system of the cognitive substrate, not by naming convention.

Construction is parameterized by three bounds. Depth bounds the longest path from the root state to any candidate leaf. Branching factor bounds the number of outbound edges admitted at any single node. Dwell time bounds the wall-clock or step-count interval during which a graph may remain in the working set before it must be either promoted, pruned, or transitioned to dormancy. The construction routine evaluates each bound at the moment of node or edge insertion. If the insertion would push the graph across any of the declared bounds, the insertion is refused and the refusal is recorded as a structural rejection event.

Out-of-bound graphs are not trimmed to fit and they are not silently truncated. They are rejected at construction, which means the working set never contains a graph whose structural properties violate the policy. This is the property that makes downstream reasoning tractable: every consumer of a planning graph can rely on the bounds as preconditions, rather than rechecking them and handling the failure case at every call site.

Lifecycle transitions are likewise governed structurally. Promotion moves a graph's selected branch into verified execution memory through an explicit transition that is itself a lineage event. Pruning removes branches that the policy has marked inadmissible; the pruned branches are retained in archival form so that forensic reconstruction can recover them, but they are no longer eligible for selection. Dormancy preserves a graph for possible future revival without keeping it in the active working set.

Operating Parameters

Each bound is declared in the policy reference and is interpreted in the context of the agent's deployment. A therapeutic agent operating under a long-horizon care plan may run with a deeper depth bound and a longer dwell time than an autonomous vehicle running a tactical maneuver, where shallow depth and short dwell are required for the planning graph to keep pace with the physical environment. The bounds are not architectural constants; they are configuration that the same architecture admits without modification.

Branching factor is typically the most sensitive parameter because it controls the combinatorial cost of evaluation. Setting it too low starves the forecasting engine of alternatives; setting it too high admits graphs whose evaluation cost dominates the inference budget. The policy reference therefore allows branching factor to be declared as a function of the depth at which the branching occurs, so that breadth at the root can differ from breadth at the frontier.

Dwell time interacts with the archival mechanism. A short dwell forces frequent promotion or dormancy events, which produces a denser archival record at the cost of more lifecycle work. A long dwell reduces lifecycle overhead but extends the interval over which a speculative graph may influence the agent's resource allocation. Operators tune dwell to match the rate at which the agent's environment is expected to invalidate prior speculation.

All three bounds are enforced at the data structure level. There is no fallback path in which a graph violating the bounds is constructed and later reconciled. This is the structural property that distinguishes the mechanism from heuristic pruning systems, which permit the violating graph to exist transiently and then attempt to recover.

Alternative Embodiments

In a tactical embodiment, depth and dwell are minimized so the forecasting engine can keep pace with rapid environmental change. The branching factor remains generous to preserve optionality at the immediate frontier. Autonomous driving and robotic manipulation are typical deployments.

In a deliberative embodiment, depth is extended and dwell is lengthened so the forecasting engine can explore long-horizon plans. Branching factor at the root is held wide while branching at deep nodes is constrained, producing a graph that is broad near the present and narrow at the horizon. Therapeutic and care-planning agents are typical deployments.

In a multi-agent embodiment, planning graphs are exchanged between cooperating agents as opaque references rather than as full structures. A receiving agent can take a producing agent's graph as the root of its own forecasting process, subject to its own bounds. The bound enforcement is per-agent, so a receiving agent will reject a graph whose imported structure would push it over its own configured limits, even if the producing agent operated under more permissive bounds.

In a constrained embodiment intended for safety-critical certification, the bounds are not only enforced but also fixed at deployment time and made non-revisable except through a formal recertification cycle. This produces an embodiment in which structural properties of the forecasting engine can be the subject of regulatory attestation, rather than only its behavioral outputs.

In an adaptive embodiment, the bounds are functions of the agent's current resource budget and observed environment dynamics. The policy reference declares the function rather than constants, and the construction routine evaluates the function at each insertion. This supports deployments that must trade exploration breadth against computational headroom in real time, without ever permitting a graph that violates the bounds in force at the moment of insertion. The adaptive embodiment is distinguished from heuristic budget-based pruning in that the bounds remain admission criteria, evaluated at construction at the moment of node insertion; the function only changes what those criteria are at each evaluation.

Composition

Planning graphs compose with the cognitive forensics archival mechanism. Because rejected constructions are recorded as lineage events, the archival record captures not only the graphs the agent considered but also those it declined to construct. This allows post-hoc analysis to distinguish "the agent did not consider option X" from "the agent attempted to consider option X and was structurally prevented."

Planning graphs compose with the inference control layer through the pre-generation distinction. A graph node may incorporate substrate-derived material, inference-generated material, or both; the source distinction is preserved at the node, so that branch evaluation can apply different admissibility rules to different provenance classes. A planning branch that depends on inference-generated content can be subjected to stricter promotion criteria than one grounded in substrate-derived content.

Planning graphs compose with policy revision. When the bounds change, in-flight graphs are re-evaluated against the new bounds at the next lifecycle transition. Graphs that satisfy the revised bounds continue; graphs that do not are pruned, with the pruning recorded. This avoids the failure mode in which a policy tightening leaves stale graphs in the working set that no longer reflect operator intent.

Rejection Semantics

A planning graph that exceeds any declared bound is not constructed. The construction routine signals a structural rejection and emits a lineage entry recording the bound that was violated, the parameter under which the violation occurred, and the input conditions that produced the attempted construction. The forecasting engine does not retry with relaxed bounds and does not silently degrade to a smaller graph. Either of those behaviors would erode the precondition that downstream consumers rely on.

Because rejections are first-class lineage events, operators can monitor the rate and distribution of rejections as a signal about whether the configured bounds match the workload. A persistent rejection rate at a particular depth suggests that the depth bound is too tight for the deployment; a persistent rejection rate at a particular branching factor suggests that the breadth allotment is mismatched against the input space. The signal is structural, not inferential, so operators do not have to design custom telemetry to extract it.

Rejection at construction is distinct from pruning during lifecycle. Pruning removes branches from a graph that was admitted; the pruned branches were structurally valid at construction and became inadmissible later under policy evaluation. Rejection prevents the graph from existing at all. The two events are recorded under distinct lineage entry types so that forensic analysis can distinguish "the agent declined to build this graph" from "the agent built this graph and then discarded part of it."

Prior Art

Classical planning systems, including STRIPS-derived planners and hierarchical task networks, structure search but do not enforce structural separation between speculative and verified state. The planner's intermediate hypothesis space and the executor's committed state share a memory model, with separation maintained by convention. The mechanism here differs in that the separation is enforced at the data structure level.

Monte Carlo Tree Search and related rollout methods bound search through computational budget but do not reject constructions structurally. A search tree that exceeds budget is truncated rather than refused, which means the consumer of the result cannot rely on uniform structural properties. The bounds described here are admission criteria, not termination criteria.

Speculative execution in processor architectures and database engines provides a precedent for structural separation between speculative and committed state, but the speculation in those systems is single-path and does not maintain a graph of alternatives that downstream reasoners can inspect. The forecasting engine generalizes the principle to a branching structure that is itself a first-class object available to subsequent stages of the cognitive pipeline.

Disclosure Scope

The disclosure covers the construction of planning graphs as data structures distinct from verified execution memory, the policy-declared bounds on depth, branching factor, and dwell time, the structural rejection of graphs that would exceed any declared bound, the lifecycle transitions of promotion, pruning, and dormancy, and the per-agent enforcement of bounds in multi-agent exchange. It covers the use of planning graphs as substrate for downstream reasoning under the precondition guarantees that bound enforcement provides. The scope extends to tactical, deliberative, multi-agent, and certification-bound embodiments, and to any cognitive architecture in which speculative planning must coexist with verified execution without contaminating it.