Mechanism
Multi-model arbitration is the configuration in which multiple probabilistic inference engines contribute candidate transitions to the same inference process. The engines share a single semantic state object and are subject to a single set of governance constraints. This is the central structural commitment: the engines do not run as independent pipelines whose results are merged afterward. They participate in one governed inference loop, against one shared record of what the inference process currently means.
At each step, one or more inference engines produce candidate transitions. Each candidate is independently mapped to a mutation descriptor and independently evaluated by the admissibility gate against the shared semantic state object. The arbitration mechanism therefore does not operate over opaque model outputs. It operates over typed, structured proposed mutations that have each been put through the same admit, reject, or decompose evaluation, which is what makes principled selection among them possible at all.
Trust-Weighted Selection Among Admitted Candidates
Arbitration is invoked when multiple candidates from different engines are admitted at the same step. The admissibility gate has already determined that each of these candidates is permissible against the shared semantic state. The arbitration engine then selects among the admitted candidates using trust-weighted evaluation, scoring each candidate according to the originating engine's trust weight, the candidate's semantic coherence with the current state, and the candidate's alignment with inference intent.
Selection therefore operates only over candidates that have already survived governance. Arbitration does not relax admissibility; a candidate that fails admissibility is never a selection option regardless of which engine produced it. The three scoring inputs (originating-engine trust weight, semantic coherence, and intent alignment) determine which of the admitted candidates advances the inference process at that step.
Dynamic Trust Weighting
The trust weight assigned to each engine is not static. An engine whose candidates are predominantly rejected accumulates negative trust-weight adjustments and is progressively de-prioritized. An engine whose candidates are predominantly admitted accumulates positive adjustments and is progressively favored. The trust weight evolves as a function of how the engine's candidates fare against the shared admissibility gate over the course of inference.
The effect is that the mechanism identifies which engines are most suitable for the current inference context without static selection heuristics. There is no fixed routing table declaring in advance which engine handles which class of work. Suitability is established empirically and continuously, by whether an engine's proposals survive admissibility against the shared semantic state, and the arbitration scoring reflects that evolving standing.
The Semantic Mutation Lifecycle
The multi-model mechanism supports a semantic mutation lifecycle in which one engine's admitted transition may be refined or qualified by a subsequent transition from a different engine. Contribution is not confined to a single decisive selection per step. An admitted transition contributed by one engine becomes part of the shared semantic state, and a later transition from another engine may elaborate, restrict, or qualify it, again subject to admissibility against the now-updated state.
This is why the engines compose into a single inference trajectory rather than producing isolated competing outputs. The semantic state carries forward the cumulative result of whichever admitted transitions, from whichever engines, advanced it. Different engines can contribute complementary pieces to the same line of inference across successive steps.
Lineage and Contribution Tracing
The semantic state object's lineage records the originating engine for each admitted transition. This recording enables contribution tracing: the inference output can be decomposed after the fact into which engine contributed which admitted transition. Because lineage already records, for each admitted transition, the mutation descriptor that was applied and the admissibility determination that permitted it, attaching the originating engine to that record makes the multi-engine provenance fully auditable.
Contribution tracing further enables identification of complementary or conflicting inference behavior across engines. The lineage shows where engines reinforced one another's transitions and where one engine's transitions repeatedly qualified or negated another's. This is a recorded, reconstructable property of the inference process, not an inferred summary statistic.
Prior-Art Distinction
Conventional multi-model systems either run models in parallel and merge their outputs by voting or weighted averaging, treating each model as an opaque output source, or route each request to a single model and discard the rest. Neither class binds the models to a shared, governed semantic state, and neither subjects every model's contribution to the same per-transition admissibility evaluation before the contribution can affect output. In those systems the only thing combined is final output. Here, the engines share one semantic state object and one set of governance constraints, each candidate is independently mapped to a mutation descriptor and independently passed through the admissibility gate, selection among admitted candidates is trust-weighted by an engine standing that evolves with the engine's admission record, and the originating engine of each admitted transition is recorded in lineage. Suitability is established by survival against governance rather than by a static routing heuristic.
Disclosure Scope
The multi-model arbitration mechanism, comprising multiple probabilistic inference engines contributing candidate transitions to a single inference process that shares one semantic state object and one set of governance constraints, the independent mapping of each candidate to a mutation descriptor and its independent evaluation by the admissibility gate, the trust-weighted selection among admitted candidates by originating-engine trust weight, semantic coherence with the current state, and alignment with inference intent, the dynamic trust weighting in which predominantly rejected engines are de-prioritized and predominantly admitted engines are favored without static selection heuristics, the semantic mutation lifecycle in which one engine's admitted transition may be refined or qualified by a subsequent transition from another engine, and the lineage recording of the originating engine for each admitted transition enabling contribution tracing and identification of complementary or conflicting inference behavior, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart) at Section 8.19. This article describes that disclosed mechanism. The scope extends to embodiments employing inference engine families not enumerated and to deployment configurations in which the shared semantic state object is realized over different representations, provided the engines remain bound to a single semantic state object and a single set of governance constraints and provided each candidate is evaluated by the same admissibility gate before arbitration.
