Executive Graph Conflict Resolution

Mechanism

The conflict-resolution mechanism activates whenever the executive graph receives two or more forecast records that share a common subject identifier, overlap in their temporal scope, and assert values that fall outside one another's reported confidence intervals. Detection is structural rather than inferential: the engine compares typed forecast records on declared fields and flags overlap-with-divergence according to a policy-defined comparator. There is no heuristic threshold to tune; the comparator is declared in the policy reference and applies uniformly.

Once a conflict is detected, the engine constructs a disagreement record. The record names each contributing source, the forecast value each source supplied, the confidence each source asserted, the method-of-derivation each source declared, and the inputs each source consumed. Crucially, the record names the locus of disagreement: it identifies whether the sources disagree on the value itself, on the confidence to attach to the value, on the temporal scope over which the value applies, on the method by which the value was produced, or on the inputs available at derivation time. The locus is computed by structural comparison of the typed forecast records, not by natural-language interpretation.

The protocol then proceeds through declared resolution phases. The first phase is independence verification: the engine examines the lineages of the conflicting forecasts to determine whether the sources are truly independent or share inputs, methods, or upstream stages. Apparent disagreement among sources that share a hidden upstream is recorded as pseudo-disagreement and treated differently from genuine disagreement among independent sources. The second phase is reconcilability assessment: the engine determines whether the disagreement is reconcilable through declared transformations (unit conversion, scope re-projection, freshness reconciliation) or whether it represents an irreducible factual divergence. Reconcilable disagreements are resolved deterministically through the declared transformations and the resolution is recorded. Irreducible disagreements proceed to the next phase.

The third phase is conditional resolution. The engine consults the policy reference for the resolution policy applicable to the subject and context. The policy may direct the engine to select the highest-confidence forecast, to select the forecast from the source with the highest declared method-strength, to refuse resolution and emit only the disagreement record, or to defer to a designated arbiter. Whatever the policy directs, the engine records both the conditionally resolved value (if any) and the disagreement record, and emits both as outputs. Downstream consumers always receive the disagreement record; the conditionally resolved value is offered as a convenience, not as a substitute for the underlying disagreement.

The disagreement record is itself a typed record carrying its own identifier, lineage, and lifecycle. It can be retrieved, audited, cited in downstream decisions, and reopened if subsequent evidence changes the picture. The engine never destroys a disagreement record by overwriting it; revisions produce new records linked to their predecessors.

Operating Parameters

The detection comparator is parameterized by tolerance, which determines how far two forecast values must diverge before they are considered in conflict given their reported confidences. Tolerance is declared per subject domain in the policy reference. Independence-verification depth determines how far back through lineage the engine traces to detect shared upstreams; deeper tracing is more thorough but more expensive. Reconcilability transformations are declared as a closed catalog: only transformations explicitly listed in the policy reference may be applied, and each transformation is itself a typed function with its own validity conditions.

Resolution-policy parameters define which resolution behavior applies under which circumstances: confidence-priority, method-priority, refuse-and-disagreement-only, or deferral-to-arbiter. Disagreement-record retention parameters govern how long records remain in the live tier before migration to historical storage. Reopening-trigger parameters define what kinds of new evidence cause a closed disagreement record to be reopened for re-evaluation.

Pseudo-disagreement classification is parameterized by an upstream-overlap threshold that determines how much shared lineage between two sources causes their apparent disagreement to be reclassified. A small shared upstream may be treated as ordinary independent disagreement; a substantial shared upstream causes the disagreement to be flagged as artifactual and routed to a different resolution path that treats the divergence as a signal about the upstream itself rather than as a substantive claim about the world. The threshold is declared per subject domain, allowing high-stakes domains to apply stricter independence requirements than low-stakes ones.

Alternative Embodiments

In a strict embodiment, the engine refuses all conditional resolution: every detected conflict produces a disagreement record and no resolved value, and downstream consumers must implement their own resolution logic. This embodiment is suitable for high-stakes domains where silent resolution is unacceptable.

In a permissive embodiment, the engine always produces a conditionally resolved value alongside the disagreement record, applying the policy's resolution rule. This embodiment is suitable for low-stakes high-throughput domains where most consumers prefer a default.

In a tiered embodiment, the resolution behavior depends on the subject's classification: high-stakes subjects produce disagreement-only outputs, low-stakes subjects produce conditionally resolved values, and intermediate subjects produce both with explicit caveats. The classification is declared per subject domain.

In an arbiter-mediated embodiment, irreducible disagreements are routed to a designated arbiter — human or automated — who records a resolution decision that becomes part of the disagreement record's lineage. The arbiter does not alter the underlying forecast records; the arbiter's decision is itself a typed output that downstream consumers can accept or reject.

In a federated embodiment, conflicting forecasts originate from engines under separate custody. The conflict-resolution mechanism operates over the federated forecast set without requiring the engines to share internal state; each engine's forecast is treated as an opaque typed record carrying its own asserted lineage, and disagreement records are exchanged across the federation as first-class artifacts.

Composition

The conflict-resolution mechanism composes with the confidence-input mechanism: per-source confidences supplied with input forecasts feed directly into the detection comparator and into confidence-priority resolution policies. It composes with the lineage mechanism: independence verification traverses the lineage records produced by the engine for each forecast. It composes with the policy-reference mechanism: the comparator, the reconcilability transformation catalog, and the resolution-policy parameters are all declared in the policy reference. It composes with downstream decision mechanisms by emitting disagreement records as typed inputs that decision logic can consume directly.

On the upstream side, the mechanism does not require the producing sources to be aware of one another. Sources continue to emit independent forecasts under their own methods; the engine assumes responsibility for detecting and structuring any disagreement that arises among them.

The mechanism further composes with identity layering on the source side: each contributing forecast's source resolves through the source's identity stack, and the disagreement record names the scope-bounded layer under which each source asserted its forecast. A disagreement among sources operating under their professional-licensure layers is recorded differently than a disagreement among sources operating under contextual-session layers, even when the numeric forecasts are identical. Auditors examining the disagreement record can therefore distinguish a substantive professional disagreement from an incidental session-level divergence without re-deriving the layer structure.

The mechanism composes with downstream decision logic by exposing the disagreement record as a structured artifact that decision rules may examine directly. A decision rule may, for example, require not merely that some forecast satisfy a confidence threshold but that no qualifying disagreement record exists for the same subject within a stated freshness window. This shifts the burden of disagreement-handling from the producing engines to the consuming decisions, which is the correct locus for the policy choice between acting on a conditionally resolved value and waiting for further evidence.

Distinction from Prior Art

Conventional ensemble and consensus systems collapse disagreement into a single output value through averaging, voting, or weighted combination. The disagreement is consumed by the aggregation and is not recoverable downstream. Even systems that report variance or confidence intervals around an aggregated output present disagreement as noise around a central tendency rather than as a structured finding about which sources claimed what under which evidence.

Multi-agent arbitration systems often elect a single authoritative outcome through voting protocols or designated arbiters, but the losing positions are typically discarded rather than preserved as typed artifacts. Truth-discovery and source-reliability frameworks model source quality but generally do so to improve aggregation, not to preserve disagreement as a first-class output. Probabilistic fusion frameworks combine distributions but do not retain a structured disagreement record naming the locus of divergence.

The disclosed mechanism differs structurally: disagreement is a typed first-class output of the engine; the locus of divergence is computed and named rather than aggregated away; conditional resolution, where it occurs, is offered alongside rather than instead of the disagreement record; and the protocol distinguishes pseudo-disagreement (shared-upstream artifacts) from genuine disagreement among independent sources. Downstream consumers always have access to the underlying disagreement and are never structurally prevented from acting on it.

Disclosure Scope

A further structural distinction concerns the treatment of disagreement over time. Conventional systems either resolve conflicts in the moment and discard the losing positions, or accumulate disagreement in unstructured logs that are difficult to query. In the disclosed mechanism, the disagreement record is a typed long-lived artifact with its own lifecycle, retrievable by subject and by participating source, and capable of being reopened when later evidence arrives. The engine therefore supports queries that conventional aggregating systems cannot answer: how often did source A and source B disagree about subjects in domain D over the last quarter, what fraction of those disagreements were reconcilable through declared transformations, and which irreducible disagreements remain open. These queries are essential for regulatory oversight of multi-source forecasting and for the accreditation of sources whose competence is itself a function of their historical agreement profile.

This article describes the conflict-resolution mechanism of the forecasting engine as disclosed in the cognition patent. Implementations in which conflicting forecasts among independent sources produce a typed disagreement record naming the locus of divergence, in which independence verification distinguishes shared-upstream pseudo-disagreement from genuine divergence, and in which conditional resolution is emitted alongside rather than in place of the disagreement record fall within the disclosed scope. The scope encompasses strict, permissive, tiered, arbiter-mediated, and federated embodiments, and is independent of the specific comparator used for detection, the specific catalog of reconcilability transformations elected by a given policy reference, and the specific transport mechanisms used to convey forecast and disagreement records.