Branch Dormancy and Deferred Promotion

Mechanism

The forecasting engine maintains a planning graph in which each node represents a candidate continuation of the agent's situation. Each branch carries a salience value derived from prior probability, expected information gain, dependency on currently active premises, and recent observational support. As salience evolves over time, branches cross thresholds that govern their participation in the active scheduler. Rather than collapse this evolution into a binary keep-or-discard event, the engine implements a three-zone lifecycle: active, dormant, and archived. Active branches are scheduled for expansion, scoring, and downstream consumption. Dormant branches are detached from the scheduler but retained in a compacted form. Archived branches are written to long-term lineage storage and are no longer eligible for in-process revival without an explicit replay operation.

Transition into dormancy is triggered when a branch's salience falls below a configured activation threshold for a configured number of evaluation cycles, or when its dependency frontier becomes inconsistent with currently verified observations. On transition, the engine performs a deterministic compaction: redundant intermediate evaluations are removed, sufficient statistics are retained, and pointers to the branch's premise set, hypothesis set, and prior posterior estimates are preserved. The compacted record is small enough that thousands of dormant branches occupy roughly the working-memory cost of a few dozen active branches, but rich enough that revival does not require recomputation from scratch.

Revival is the inverse operation. The dormancy manager subscribes to a set of revival predicates derived from observed mesh state: a previously contradicted premise becoming consistent again, the appearance of a salient observation that matches a dormant branch's signature, or a downstream policy explicitly requesting reconsideration of an alternative path. When a predicate fires, the dormant branch is rehydrated into the active region with its preserved sufficient statistics; the scheduler then reincorporates it into the next evaluation cycle. Because the lineage was preserved, the rehydrated branch begins from its prior posterior rather than from a flat prior, which is the property that makes deferred promotion materially cheaper than rebuilding the branch from observation history.

Pruning, in this architecture, refers only to the active-to-dormant transition or to the dormant-to-archived transition. The engine never silently discards a branch that has contributed to any committed downstream artifact. The audit lineage is monotonic with respect to creation events; only the operational status of a branch changes, and every status change is itself a recorded event. This separation between operational status and existential record is what allows the engine to explore aggressively without losing the ability to explain why a previously considered alternative was set aside.

Operating Parameters

A deployment of the dormancy manager is governed by a small set of declarative parameters. The activation threshold sets the salience floor below which a branch becomes a candidate for dormancy; the dwell window specifies how many consecutive cycles a branch must remain below the threshold before transition occurs, which damps oscillation around the boundary. The dormant capacity bound caps the number of dormant records the engine will retain; when the bound is reached, the lowest-salience dormant branches are promoted to archive in policy-ranked order. The revival predicate set defines which mesh-state changes are eligible to trigger rehydration, and a per-revival cost ceiling prevents pathological cascades in which a single observation rehydrates an unbounded population of dormant branches simultaneously.

The salience function itself is treated as a parameter rather than a fixed component. Operators may select among posterior-weighted, information-gain-weighted, and consequence-weighted salience definitions, and may compose them through declared mixing weights. Each salience definition is recorded in the policy reference and is referenced by hash from every transition record, so that an audit can reconstruct exactly which definition governed a given branch's lifecycle at each point in time. Two further parameters govern the integrity of the lifecycle. The compaction profile specifies which sufficient statistics are retained on dormancy transition; profiles range from minimal, retaining only the posterior summary, to verbose, retaining the full evaluation trace. The archive cadence determines how frequently dormant records below a configured residual-salience floor are promoted to long-term storage. Together these parameters are tuned to bound working-memory cost while preserving the revival economics that justify dormancy in the first place.

Alternative Embodiments

The dormancy manager admits several embodiments that vary the mechanics without changing the structural contract. In a tier-graded embodiment, the binary active-versus-dormant distinction is generalized to N tiers, each with its own scheduler weight and compaction profile, so that a branch decays gracefully through partial-attention states before fully detaching. In a content-addressed embodiment, dormant branches are indexed by a hash of their premise signature, allowing observations to query the dormant set in constant time and rehydrate matches without scanning. In a federated embodiment, dormant branches generated by one agent are advertised to peers via the mesh substrate, enabling another agent that encounters a relevant signature to revive a peer's prior reasoning rather than recompute it. A revival-by-counterfactual embodiment further allows policy modules to inject hypothetical observations to probe the dormant set, which is useful for stress-testing and what-if analysis without disturbing the live scheduler.

Additional embodiments address resource-constrained and adversarial environments. A memory-tiered embodiment places active branches in fast in-process memory, dormant branches in a slower in-process region with explicit eviction barriers, and archived branches in external durable storage; the engine performs deterministic page-in operations on revival and treats page-in latency as a first-class scheduler input. An encrypted-at-rest embodiment encrypts dormant records under a key that is itself anchored in the policy reference, which means a dormant branch cannot be silently introspected by code that does not hold the policy-bound key. A capability-scoped embodiment further partitions the dormant set into capability domains, so that an agent operating under a restricted role can revive only the dormant branches its role permits, a property useful in multi-tenant deployments and in regulated domains where exploration history is subject to compartmentalization.

A streaming embodiment pipes dormancy and revival events to an external observability channel without copying branch contents, so that an operator dashboard can surface revival rate, dwell distribution, and archive throughput as continuously updated signals. A learned-priority embodiment replaces the static activation threshold with a policy-bound model that proposes salience adjustments, while still subjecting every transition to the deterministic threshold gate, so the learning component cannot bypass the structural lifecycle. These embodiments illustrate that the structural contract, namely a bounded dormant population, preserved sufficient statistics, predicate-driven revival, and monotonic lineage, is independent of the choice of indexing strategy, storage tier, or scoring source.

Composition with Other Cognition Mechanisms

Branch dormancy composes directly with temporal anchoring: a dormant branch retains the mesh-time anchors of its last evaluation, and on revival those anchors are revalidated before the branch is allowed to influence downstream policy. If anchor revalidation fails, the branch is archived rather than rehydrated, ensuring that stale forecasts cannot leak into current operation. Dormancy also composes with executive aggregation: dormant branches are excluded from the executive summary by default but may be referenced by lineage identifier, so that a downstream consumer auditing a decision can recover the reasoning the agent considered and set aside. Together with the planning-graph containment boundary, dormancy makes the engine's exploration history a first-class, queryable artifact rather than a transient log.

Distinction from Prior Art

Conventional planning systems treat low-salience branches as garbage, discarding them when memory pressure or beam-width limits force a contraction. This produces forecasting systems that are confident in the moment but cannot explain why a particular alternative was abandoned, and that must rebuild abandoned reasoning from scratch when conditions change. Cache-based systems retain prior computations but key them by input rather than by hypothesis, which means they can reuse a calculation but cannot revive a structured chain of inferences. Probabilistic programming runtimes offer trace replay, but typically require deterministic re-execution rather than constant-time rehydration. The dormancy manager is distinct in three respects: it preserves hypothesis-level structure rather than merely cached values, it bounds the dormant population through declarative policy rather than heuristic eviction, and it ties revival to mesh-state predicates rather than to repeated input matching.

Disclosure Scope

This disclosure covers any forecasting subsystem in which branches transition between active, dormant, and archived states under declarative policy; in which dormant branches retain sufficient statistics enabling revival at lower cost than recomputation; in which revival is triggered by observable mesh-state predicates; and in which all status transitions are recorded in a monotonic audit lineage. The mechanism is applicable across deployment domains including autonomous vehicles, companion agents, therapeutic systems, and enterprise planning, and is not limited to any particular probabilistic representation, scheduler implementation, or compaction profile.

The disclosure further encompasses systems in which the dormancy boundary is enforced at the data structure level, as opposed to systems in which it exists only as a runtime convention. It encompasses systems in which the population of dormant branches is bounded by a declared capacity rather than by ad hoc memory pressure heuristics, and systems in which branch revival is gated by predicates evaluated against a verifiable observation substrate rather than by free-running pattern matchers. The mechanism is independent of the underlying scheduler discipline and may be combined with first-fit, priority-queue, or beam-style schedulers without altering its structural contract.

The disclosure also extends to operational regimes in which the dormancy manager itself is reconfigured at runtime through declarative policy updates, with each policy revision recorded in the same monotonic lineage as the branches it governs. A reconfiguration that tightens the activation threshold or shrinks the dormant capacity bound is applied prospectively to new transitions and applied retroactively to existing dormant branches only through a recorded sweep operation, preserving the property that every status change is explained by a recorded event. The mechanism is therefore suitable for deployments that must demonstrate continuous compliance with evolving governance constraints, including domains where regulators may require evidence that no relevant alternative was discarded without a recorded justification.