Abstract
This document discloses, in enabling detail, a method and system for selecting, on a per-turn basis, which large language model (LLM) answers a given request in a multi-model deployment. The distinguishing property is that capability feasibility is a hard gate evaluated strictly before any cost or difficulty optimization, and a budget is structurally unable to override a capability-forced selection. A four-stage task assessor emits a set of required capability primitives (tooluse, vision, longcontext) under a monotone taint invariant: each later stage may raise, but never lower, a hard requirement detected by an earlier stage, while websearch is deliberately kept soft so that a need for fresh information never forces an up-route. A five-step selector then (1) filters the candidate ladder to the feasible set, (2) applies a per-task SLA floor, (3) health-gates an alpha-blended cheapest-by-quality argmax, (4) applies a low-confidence "quality bump" that is disabled whenever a Differential Item Functioning (DIF) fairness harness is paused, and (5) resolves seat-allowlist conflicts by substituting the cheapest allowed model meeting the SLA floor rather than the ceiling. Length is measured in tokens rather than characters to equalize routing cost across language cohorts. Eval/judge routes bypass the router entirely (fail-closed). A clean-room, dependency-free reference implementation is included.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Assuncao, gustavo matthew, "Capability-Gate-First Cost-Optimal LLM Task Router with Monotone Hard-Capability Taint, DIF-Fairness-Governed Escalation, and Cheapest-Allowed Self-Clamp", Technical Disclosure Commons, (July 13, 2026)
https://www.tdcommons.org/dpubs_series/10870