Cascade Forecasting in the Planning Graph

Mechanism

Cascade forecasting is the forecasting engine's ability to project a chain of speculative mutations forward through more than one hypothetical step before the agent commits to any of them. The unit of this projection is the planning graph: a mutable, memory-referenced, directed structure whose root node is the agent's current verified state and whose branches each represent a distinct hypothetical trajectory. A branch is not a single guess about the next action. It is a speculative mutation sequence, a chain of hypothetical state transitions, delegation outcomes, environmental transitions, or intent resolutions, terminating in a projected outcome that characterizes the expected terminal state of the branch if it were executed. The agent reasons about a cascade of future steps, not just the immediate one, and it does so in a computational domain that is structurally separated from its verified execution memory.

This separation is the load-bearing property. The planning graph domain and the verified execution memory domain are distinct, and no speculative branch can become verified state except by passing through a single governance-validated promotion interface. Constructing a branch that projects a successful outcome does not give the agent that outcome: the projection remains speculative until promoted and executed. Because the domains are separated, the agent can hold multiple contradictory cascades at once, one branch projecting success and another projecting failure, without producing any inconsistency in its verified state. The forecasting engine builds and evaluates the cascade; the promotion interface, not the forecast, decides what becomes real.

Speculative Mutation Simulation

For each active branch, the forecasting engine simulates the hypothetical mutations the branch represents by applying the speculative mutation sequence to a sandboxed copy of the agent's state, never to verified execution memory, and computing the branch's projected outcome. The simulation includes projected environmental responses, projected delegation outcomes where the branch involves delegation, and projected secondary effects on the agent's affective state and integrity field. The simulation is deterministic: given the same input state and the same mutation sequence, it produces the same projected outcome.

The specification distinguishes this deterministic, structurally constrained simulation from statistical tree search methods such as Monte Carlo Tree Search. Each simulation step operates deterministically on defined structural fields, produces reproducible projected outcomes, and is constrained by trust slope continuity and policy compatibility at every step, rather than being scored statistically over random rollouts. A cascade here is a defined sequence of field transitions, not a sampled distribution of playouts.

Slope-Constrained Cascades

The cascade is slope-constrained. The trust slope trajectory acts as a hard structural filter on which hypothetical futures the agent is even permitted to evaluate for promotion. For each speculative branch, the slope validation module computes a hypothetical Derived Anchor Hash, written DAH-prime, by applying the branch's speculative mutation sequence to a sandboxed copy of the agent's lineage and computing the trust slope hash that would result. The hypothetical DAH-prime is then compared against the agent's current trust slope trajectory using the same continuity validation algorithm the governance infrastructure applies to committed mutations. If the hash chain relationship between the current DAH and the hypothetical DAH-prime satisfies the cryptographic lineage requirements, the branch is slope-eligible; if the speculative mutations would produce a lineage gap, a hash chain discontinuity, or a provenance violation, the branch is slope-ineligible.

This filtering is prospective. The slope constraint screens speculative branches before they reach the promotion interface, so the governance pipeline never receives a promotion candidate that would fail trust slope validation. The specification notes this prospective filtering is more efficient than constructing every possible branch and rejecting it at promotion time, and it concentrates the agent's cognitive resources on branches that have a viable path to execution. Only slope-eligible branches may be promoted. A slope-ineligible branch may be retained for introspection, so the agent can understand why certain futures are structurally foreclosed, but it cannot advance through the promotion interface.

The Forecasting Execution Cycle

The cascade is built and evaluated by a six-phase execution cycle that runs synchronously whenever the agent faces a decision requiring speculative evaluation. Phase one, initialization, reads the agent's current verified state into a root node and either refreshes an existing planning graph or generates an initial branch set. Phase two, speculative mutation simulation, projects each branch forward as described above. Phase three, slope projection and validation, computes the hypothetical DAH-prime for each simulated branch and confirms whether it is slope-eligible, also computing the magnitude of trust slope continuation so eligible branches can be compared. Phase four, policy compatibility check, evaluates each slope-eligible branch against the agent's current policy configuration and flags branches containing policy-excluded mutations. Phase five, emotional reinforcement tagging, assigns each surviving branch an affective reinforcement tag encoding how well its projected outcome aligns with the agent's current affective disposition. Phase six, branch marking and pruning, assigns each branch a classification label and schedules failed branches for removal.

The phases run in sequence, and the cascade only advances as far as the evaluation permits. A branch that fails slope validation or policy compatibility is marked pruned before it can extend further. The cycle re-runs at each cognitive decision point, so the cascade is continuously re-projected against the agent's current verified state rather than computed once and trusted thereafter.

Branch Classification

Every branch in the cascade carries one of four classification labels. An eligible branch has passed slope validation, satisfied policy compatibility, and received positive or neutral affective reinforcement; it is a viable promotion candidate and is ranked against other eligible branches by a composite score combining projected outcome quality, trust slope continuation magnitude, integrity impact projection, affective reinforcement strength, and alignment with the agent's current intent. An introspective branch has passed slope and policy validation but received negative affective reinforcement; it is retained, not promoted, so the agent can examine why a structurally viable future is aversive to it. A delegable branch is slope-eligible and policy-compatible but better suited for transfer to a child agent. A pruned branch has failed slope validation, failed policy compatibility, exceeded an entropy or compute threshold, or been superseded by a higher-ranked branch.

Classification is not permanent. The specification states that an introspective branch may become eligible if the agent's affective state shifts, an eligible branch may become pruned if the conditions that supported it change, and a delegable branch may become eligible if its delegation target is unavailable. Because the forecasting execution cycle re-evaluates classifications each iteration, the cascade reflects the agent's current cognitive landscape rather than a stale snapshot.

Temporal Anchoring and Pruning

Each branch in the cascade is temporally anchored: it carries a timestamp recording the verified state from which it was generated and a projection window specifying the temporal range over which its speculative projections are considered valid. The projection window's duration is determined by the agent's temporal planning horizon trait and the policy configuration. Temporal anchoring prevents branches from persisting indefinitely and becoming stale as the agent's verified state and environment move past the assumptions that informed them.

The pruning manager enforces several criteria that govern the lifecycle of cascade branches: temporal expiration when a branch's projection window passes, slope invalidation when a committed mutation renders a branch's hypothetical DAH-prime discontinuous, policy revocation when a policy change makes a branch's mutations inadmissible, entropy threshold pruning when the active branch count exceeds a policy-defined limit, compute budget pruning when the cost of maintaining the cascade exceeds budget, and mutation-triggered pruning when a significant verified-state change invalidates a branch's root assumptions. Pruning events are recorded in the agent's lineage as cognitive metadata, capturing which branches were pruned, the criterion that triggered removal, and the branch's evaluation state at the time. The cascade is therefore bounded: it does not expand without limit, and its lifecycle is auditable.

Forecasting-Modulated Discovery Traversal

The cascade projection is what lets a discovery object traverse the adaptive index strategically rather than greedily. At each anchor node during traversal, the discovery object's forecasting engine constructs a planning graph in which each branch represents a candidate transition to a different neighboring anchor. For each candidate, the engine simulates the projected outcome of the transition, the projected state of the discovery object afterward, the projected semantic neighborhood that would become accessible, and the projected proximity of the post-transition state to the object's intent, then runs the full forecasting execution cycle over those candidates.

By simulating multi-step traversal sequences as planning graph branches, the discovery object evaluates not just the immediate next anchor but the projected trajectory of several future traversal steps. It can therefore identify paths that look suboptimal at the next step yet lead to superior outcomes over a longer horizon, and avoid paths that look promising at the next step yet lead to dead ends or policy violations within the projection window. The specification describes this as transforming traversal from a greedy, step-by-step process into a strategically guided one that accounts for the structure of the semantic landscape beyond the immediately visible neighborhood. The discovery object's affective state modulates the cascade: elevated risk sensitivity favors conservative paths in well-characterized neighborhoods, while elevated novelty appetite explores less-traversed neighborhoods.

Disclosure Scope

Cascade forecasting, comprising the planning graph as a first-class cognitive structure separated from verified execution memory, the projection of speculative mutation sequences forward across a projection window, the deterministic slope-constrained speculative simulation that computes a hypothetical Derived Anchor Hash and admits only slope-eligible branches, the six-phase forecasting execution cycle, the eligible, introspective, delegable, and pruned branch classification taxonomy, the temporal anchoring and pruning-manager lifecycle, and the forecasting-modulated discovery traversal that projects multi-step traversal sequences, is disclosed in United States Patent Application 19/647,395. This article describes that disclosed mechanism and does not introduce mechanisms, terminology, or quantities beyond it. The scope extends to embodiments in which the cascade is projected over different field representations and to the substrate deployment models described in the specification, provided the speculative domain remains structurally separated from verified execution memory and every promotion passes through the governance-validated promotion interface.