Executive Engine Multi-Agent Graph Aggregation

Mechanism

The executive engine sits downstream of N forecasting streams. Each stream emits a planning graph whose nodes represent candidate continuations of a shared situation, annotated with salience, evidence support, mesh-time anchors, and a dependency frontier. Streams may differ in temporal horizon, hypothesis class, or operating role; for example, one stream may forecast environmental dynamics while another forecasts the agent's own commitments. The executive engine consumes these graphs concurrently and produces a single executive summary: a bounded planning graph whose nodes correspond to coherent forecasts that downstream policy modules can act upon.

Aggregation proceeds in three phases. The first phase is alignment. Branches across streams that refer to the same underlying event class, the same mesh-time interval, and a compatible premise set are grouped into alignment clusters using a deterministic equivalence test rather than a learned similarity score. Branches that fail to align with any cluster either form new singleton clusters or are recorded as orphan candidates depending on policy. The alignment test is reproducible: given the same set of input graphs, the cluster assignment is bit-identical across runs.

The second phase is arbitration. Within each alignment cluster, the engine combines the contributing branches into a single representative node using weighted arbitration. Weights are derived from declared stream credibilities, the evidentiary support of each contributing branch, and any conflict markers introduced by the streams themselves. The combination operator is policy-selectable: a default operator computes a posterior mixture, while alternative operators implement consensus, worst-case, or veto-respecting semantics. When contributing branches disagree beyond a configured tolerance, the engine emits a conflict record that is included in the summary as a first-class artifact rather than silently smoothed away. Downstream policy can then read either the aggregated forecast, the conflict record, or both.

The third phase is compression. The aggregated graph is reduced to fit a declared bounded summary budget, expressed as a maximum number of nodes, a maximum cumulative description length, or both. Compression is performed by a policy-defined ordering function that selects which nodes survive at full fidelity, which are merged with sibling nodes under explicit subsumption rules, and which are demoted to lineage references that downstream consumers can dereference if needed. At every step, each surviving summary node retains a manifest of the contributing branch identifiers and the arbitration weights used; this manifest is what makes the audit lineage from summary back to source streams complete.

Operating Parameters

A deployment is governed by several declarative parameters. Stream credibility weights assign a base trust level to each contributing forecasting stream and may be conditioned on the operating regime. The alignment tolerance specifies how strictly two branches must match on event class, mesh-time interval, and premise set to be grouped. The arbitration operator selects the combination semantics for each cluster. The conflict threshold sets the divergence above which an explicit conflict record is emitted rather than a silent merge. The summary budget caps the size of the executive graph in nodes and description length, and the demotion policy defines the order in which nodes lose fidelity when the budget is exceeded.

Additional parameters refine the alignment phase. The event-class taxonomy determines which symbolic categories two branches must share to be considered for grouping, and may be domain-tuned so that, for example, two branches forecasting different categories of pedestrian behavior are grouped only at the appropriate level of abstraction. The premise compatibility predicate determines what counts as a compatible premise set; strict compatibility requires identical premises, while relaxed compatibility allows premises that are pairwise consistent under the engine's logical model. The mesh-time interval test specifies how strictly the temporal coverage of two branches must overlap, balancing alignment recall against alignment precision. Two further parameters govern auditability. The lineage retention profile specifies whether the engine retains pointers to contributing branch identifiers, full premise sets, or full evaluation traces; richer profiles support deeper audit at the cost of summary size. The signature scheme specifies how each summary node is bound to its contributing branches in a tamper-evident way, so that a downstream consumer can verify that a summary node was actually constituted from the branches its manifest claims.

Alternative Embodiments

The aggregation mechanism admits several embodiments. A flat embodiment runs alignment, arbitration, and compression as a single pass over the union of input graphs and is appropriate when the contributing streams are roughly co-temporal. A staged embodiment performs alignment incrementally as streams emit updates, maintaining a continuously valid summary at the cost of more complex bookkeeping. A hierarchical embodiment groups streams into tiers; tier outputs are aggregated and the resulting tier summaries are themselves aggregated, which is useful when streams come from heterogeneous subsystems with different update cadences. A federated embodiment exposes the executive summary across the mesh substrate so that peer agents may consume aggregated forecasts without requiring access to the underlying contributing streams. In each embodiment, the contract is identical: the summary is bounded, lineage is preserved, and conflicts are explicit.

A speculative-rollback embodiment maintains a tentative executive summary that may be revised when late-arriving stream updates change the alignment outcome; the previous summary version is retained in lineage so that any downstream policy that already read it can be informed of the supersession. A capability-scoped embodiment partitions streams by trust domain and produces multiple summaries, each accessible to a distinct downstream consumer, while maintaining a single underlying alignment graph; this is useful when an executive must produce different views to consumers with different need-to-know boundaries. A learning-augmented embodiment uses a policy-bound model to propose alignment cluster boundaries that the deterministic equivalence test then ratifies or rejects, allowing fast heuristics to drive throughput while structural correctness remains anchored to the deterministic check.

A streaming embodiment emits the executive summary as a continuously updated artifact rather than as a single snapshot, with each update accompanied by a delta describing which clusters were added, modified, or retired since the last emission. A pull embodiment inverts the data flow so that downstream policy describes the regions of forecast space it cares about and the engine returns aggregations bounded to those regions, which is useful when the downstream consumer has stronger constraints on its reading cost than the engine has on its production cost. Across all embodiments, the same five contracts hold: alignment is deterministic, arbitration weights are derived from declared sources, conflicts above tolerance are first-class artifacts, the summary is bounded by a declared budget, and the lineage from each summary node back to its contributing branches is complete and verifiable.

Composition with Other Cognition Mechanisms

A further composition concern is the relationship between aggregation cycles and downstream commitment. Because the executive summary is the canonical input to commitment policy, its update cadence interacts with the agent's overall reaction latency. The engine therefore supports a commitment-aware mode in which a partial aggregation may be exposed to downstream policy with an explicit completeness flag, allowing a time-critical consumer to act on a known-incomplete summary while a non-critical consumer waits for the completed pass. The completeness flag is itself an audit artifact, so a later review can determine which commitments were taken against partial summaries and which against complete ones. Executive aggregation composes with temporal anchoring by carrying anchor sets through the alignment and arbitration phases; a summary node inherits the union of its contributing anchors and is invalidated whenever any contributing anchor is repudiated. It composes with branch dormancy by treating dormant branches as eligible inputs only on explicit revival; a dormant branch never silently appears in the summary. It composes with the planning-graph containment boundary by ensuring that the executive summary itself is treated as forecasting state until a downstream policy explicitly commits an action, at which point only the committed action crosses into verified execution memory. The conflict records emitted during arbitration are themselves auditable artifacts that compose with downstream governance modules, which may choose to halt commitment, request human review, or invoke fallback policies when a conflict exceeds a domain-specific severity.

Distinction from Prior Art

Existing multi-source forecast aggregation typically falls into one of three families. Ensemble averaging combines numerical predictions but loses structural information about why each predictor reached its conclusion. Voting and arbitration schemes preserve structure but generally operate on flat decision sets rather than on graph-structured forecasts with shared lineage. Knowledge-graph fusion methods merge structured outputs but rarely impose a bounded summary budget, with the result that aggregation is unbounded in cost and the downstream consumer must implement its own summarization. The executive engine is distinct in combining graph-aware alignment, weighted arbitration with explicit conflict records, bounded compression under declared budgets, and end-to-end audit lineage in a single mechanism. The bounded compression property is structurally significant: it allows downstream policy to commit to a fixed reading cost regardless of how many or how prolific the contributing streams become.

Disclosure Scope

This disclosure covers any subsystem in which two or more forecasting streams produce graph-structured outputs that are aggregated into a single bounded summary; in which alignment, arbitration, and compression are governed by declarative policy; in which conflict between contributing branches is preserved as an explicit summary artifact; and in which an audit lineage from each surviving summary node back to its contributing branches is retained. The mechanism is applicable across deployment domains and is not limited to any particular arbitration operator, compression algorithm, or stream representation.

The disclosure further encompasses systems in which the summary budget is enforced as a hard cap rather than a soft target, systems in which the conflict tolerance is itself a declared parameter rather than an implementation constant, and systems in which the alignment, arbitration, and compression phases are independently observable and independently auditable. It encompasses deployments in which the executive summary is produced once per cycle, deployments in which it is maintained continuously, and deployments in which it is produced on demand against a snapshot of the contributing streams. The mechanism is also applicable in regulated environments where the lineage manifest is required to be presented as evidence; in such environments the manifest serves not only operational reproducibility but also satisfies external auditors that no contributing branch has been silently dropped, reweighted, or merged outside the policy-declared rules.