SAFE-LOOP: Detecting Silent Sensor Faults in LLM Control Agents with a Contract Gate

SAFE-LOOP is a contract gate for LLM control agents. LLM agents now drive actuators unattended, so a stale reading becomes a silent wrong action. The gate sits between agent and actuators and builds its own belief from exactly the agent's readings. Over three simulated testbeds and one ICU replay, it detected 89.6–92.9% of stale readings on every channel tested; wrong outputs inside the drift window at 14.8% (HVAC) and 35.2% (wearable) with the fixed drift bound and 90.1% with a fitted HVAC drift model on four open models, and the ICU replay showed no measurable gain on them. With stale HVAC readings the gate cut the silent-fault rate from 99% to 10% (CI 6 to 15) with Claude Haiku 5.5 and from 53% to 7% (CI 3 to 12) with Claude Opus 5.5. Pooled over nine models, it removed 0.90 stale-reading HVAC violations per episode (CI 0.30 to 1.51) at 1.03–1.06 times the episode cost. On greenhouse moisture a per-step exposure bound holds up to 0.65 from the set x ≤ 0.58; its recovery hypothesis is checked on logs, not certified. The gate refuses 23.7% of fault-free HVAC reads (qwen3-235b) and locks out 31.3% of no-fault wearable decisions. Preprint. Code and data: 10.5281/zenodo.24186989.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-13
DOI
https://doi.org/10.5281/zenodo.24150266
Citations
1
Primary Topic
Fault Detection and Control Systems
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

SAFE-LOOP: Detecting Silent Sensor Faults in LLM Control Agents with a Contract Gate

Tharun Reddy Kandukuri
1 citations
Zenodo (CERN European Organization for Nuclear Research)
Fault Detection and Control Systems
preprint

SAFE-LOOP: Detecting Silent Sensor Faults in LLM Control Agents with a Contract Gate

Tharun Reddy Kandukuri
preprint en
1 citations

Abstract

SAFE-LOOP is a contract gate for LLM control agents. LLM agents now drive actuators unattended, so a stale reading becomes a silent wrong action. The gate sits between agent and actuators and builds its own belief from exactly the agent's readings. Over three simulated testbeds and one ICU replay, it detected 89.6–92.9% of stale readings on every channel tested; wrong outputs inside the drift window at 14.8% (HVAC) and 35.2% (wearable) with the fixed drift bound and 90.1% with a fitted HVAC drift model on four open models, and the ICU replay showed no measurable gain on them. With stale HVAC readings the gate cut the silent-fault rate from 99% to 10% (CI 6 to 15) with Claude Haiku 5.5 and from 53% to 7% (CI 3 to 12) with Claude Opus 5.5. Pooled over nine models, it removed 0.90 stale-reading HVAC violations per episode (CI 0.30 to 1.51) at 1.03–1.06 times the episode cost. On greenhouse moisture a per-step exposure bound holds up to 0.65 from the set x ≤ 0.58; its recovery hypothesis is checked on logs, not certified. The gate refuses 23.7% of fault-free HVAC reads (qwen3-235b) and locks out 31.3% of no-fault wearable decisions. Preprint. Code and data: 10.5281/zenodo.24186989.

Zenodo (CERN European Organization for Nuclear Research)
Fault Detection and Control Systems
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.

Laika access coming soon