INV-15: A Formal Safety Invariant for KV-Cache Reuse in Multi-Agent Judge Pipelines

Cross-agent KV-cache reuse delivers 7-17x compression in multi-agent LLM pipelines, but silently corrupts judge/critic agents via the Judge Candidate Reuse (JCR) failure mode. We introduce INV-15, a formal safety invariant requiring judge-class agents with JCR risk score above 0.7 to use dense prefill. Implemented as the JCR Safety Gate in ContextForge (AMD Instinct MI300X, ROCm 7.x, vLLM V1 ATOM plugin). Results: 0 INV-15 violations, critic dense prefill rate 1.000, compatible with 10.81x TokenDance compression. First production implementation of a formal safety invariant for cross-agent KV-cache reuse.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-05-18
DOI
https://doi.org/10.5281/zenodo.20277875
Primary Topic
Distributed systems and fault tolerance
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
preprint

INV-15: A Formal Safety Invariant for KV-Cache Reuse in Multi-Agent Judge Pipelines

Pablo M. Suarez
Zenodo (CERN European Organization for Nuclear Research)
Distributed systems and fault tolerance
preprint

INV-15: A Formal Safety Invariant for KV-Cache Reuse in Multi-Agent Judge Pipelines

Pablo M. Suarez
preprint en

Abstract

Cross-agent KV-cache reuse delivers 7-17x compression in multi-agent LLM pipelines, but silently corrupts judge/critic agents via the Judge Candidate Reuse (JCR) failure mode. We introduce INV-15, a formal safety invariant requiring judge-class agents with JCR risk score above 0.7 to use dense prefill. Implemented as the JCR Safety Gate in ContextForge (AMD Instinct MI300X, ROCm 7.x, vLLM V1 ATOM plugin). Results: 0 INV-15 violations, critic dense prefill rate 1.000, compatible with 10.81x TokenDance compression. First production implementation of a formal safety invariant for cross-agent KV-cache reuse.

Zenodo (CERN European Organization for Nuclear Research)
Peace, Justice and strong institutions
Distributed systems and fault tolerance
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.