What question did this study set out to answer?

The aim is to develop a compute-first safety framework for LLM routing that operates without privileged meta-judges.

February 8, 2026Open Access

Compute-First Safe LLM Routing Without Meta-Judges Observable Viability with Information and Energy Budgets

Key Points

The aim is to develop a compute-first safety framework for LLM routing that operates without privileged meta-judges.
Formalizes observable viability under information and energy budgets.
Synthesizes policies offline to minimize information rate while satisfying constraints.
Develops the Rate–Viability Function (RVF) and Energy–Viability Frontier (EVF).
Specifies tamper-evident logging and audit dynamics.
Provides a layered rollout path and implementation templates for AI engineers.
Establishes an effective framework for minimizing information rate and resource use.
Introduces new theoretical concepts like the RVF and EVF for optimal routing.
Demonstrates robust performance under various conditions and constraints.

Abstract

This preprint introduces a compute-first safety framework for LLM routing that works without privileged meta-judges. The method formalizes observable viability under explicit information/compute and energy budgets, using a finite-dimensional rational-inattention objective plus a hardware-grounded energy calibration layer. Policies are synthesized offline to minimize information rate while satisfying auditable constraints on violation risk, compute, energy, and optional service/utility floors, then compiled into lightweight runtime tables for fail-closed deployment.The paper develops the theory of a Rate–Viability Function (RVF) and an Energy–Viability Frontier (EVF), including dual/KKT characterization, exponential-form optimizers, monotone budget tuning, and robustification for surrogate misspecification, OOD drift, delayed feedback, and risk-debt stabilization. It also specifies tamper-evident logging and retroactive audit/slashing dynamics to correct underestimation risk in no-meta settings.To reduce adoption friction for AI engineers, the manuscript provides a layered rollout path (L0–L4) and implementation-ready project templates in /configs, /schemas, and /scripts (YAML/JSON/Python), enabling immediate integration into production LLM agent pipelines where token, latency, and Joule efficiency are first-class constraints.

Ask AI

Helpful

Bookmark

View Full Paper