A Predictive Safety Execution Kernel for External Planners: System Implementation and Independent Recomputation under a Frozen Protocol

Hengjie is a runtime quality-assurance mechanism for control software driven by external planners, including large language models (LLMs). Because planner output is untrusted, release is gated: the kernel releases a candidate only after robust predictive verification against state uncertainty and the committed queue, with a certified fallback. For the discrete-time model, every released action remains covered by the execution invariant under the premises, for any proposer. The characterisation is a pre-registered, frozen-protocol study of a simulated thermal process under bounded faults (424 executable cells, failures and counterexamples retained). The protected configurations (legacy B, 34 cells; B2, 88 cells) recorded 0 violation minutes and 0 uncovered-authorisation minutes; the unprotected configuration recorded 25,792 and 34,433, respectively. Under one-slot delay the protected paths stayed safe but completed 10–11 of 1,440 task minutes. A pre-registered budget study re-ran the reviewed LLM cells at 4× and 16× the frozen budget and cut fallback share from 0.964 to 0.507, but left the task ratio identical (0.9981; non-LLM configuration 0.9983): budget changed which controller holds the band, not the outcome. A control-barrier-function safety filter wrapping the same proposer, run as an exact replica in the same slots as a twelve-cell engineering baseline, recorded zero violation minutes and held the disconnection cells where the heuristic chain B collapsed. Decision latency: median 1.32–1.37 s per 60 s tick. All frozen-batch quantities were independently recomputed from a public replication package with zero discrepancies; reproducing a published predictive-safety baseline deviated by +31.4%, the cause undetermined after pre-registered single-factor diagnosis.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11778674/v1
Primary Topic
Formal Methods in Verification
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Predictive Safety Execution Kernel for External Planners: System Implementation and Independent Recomputation under a Frozen Protocol

Jiwei Wang, Kaihang Ma
Research Square
Formal Methods in Verification
preprint

A Predictive Safety Execution Kernel for External Planners: System Implementation and Independent Recomputation under a Frozen Protocol

Jiwei Wang, Kaihang Ma
preprint en

Abstract

Hengjie is a runtime quality-assurance mechanism for control software driven by external planners, including large language models (LLMs). Because planner output is untrusted, release is gated: the kernel releases a candidate only after robust predictive verification against state uncertainty and the committed queue, with a certified fallback. For the discrete-time model, every released action remains covered by the execution invariant under the premises, for any proposer. The characterisation is a pre-registered, frozen-protocol study of a simulated thermal process under bounded faults (424 executable cells, failures and counterexamples retained). The protected configurations (legacy B, 34 cells; B2, 88 cells) recorded 0 violation minutes and 0 uncovered-authorisation minutes; the unprotected configuration recorded 25,792 and 34,433, respectively. Under one-slot delay the protected paths stayed safe but completed 10–11 of 1,440 task minutes. A pre-registered budget study re-ran the reviewed LLM cells at 4× and 16× the frozen budget and cut fallback share from 0.964 to 0.507, but left the task ratio identical (0.9981; non-LLM configuration 0.9983): budget changed which controller holds the band, not the outcome. A control-barrier-function safety filter wrapping the same proposer, run as an exact replica in the same slots as a twelve-cell engineering baseline, recorded zero violation minutes and held the disconnection cells where the heuristic chain B collapsed. Decision latency: median 1.32–1.37 s per 60 s tick. All frozen-batch quantities were independently recomputed from a public replication package with zero discrepancies; reproducing a published predictive-safety baseline deviated by +31.4%, the cause undetermined after pre-registered single-factor diagnosis.

Research Square
Yanching Institute of Technology (CN)
Formal Methods in Verification
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