Confidence-Gated Neural Governance for Autonomous Software Systems: A Hybrid Deterministic-Neural Architecture with Anomaly-Aware Safety Braking

Read full text Summarise with VIRAT
Authors: Rafal Chalupka
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

We present a confidence-gated neural governance architecture for autonomous software systems that combines a deterministic decision kernel with an evolving LSTM neural layer. The system makes eight classes of governance decisions based on continuous monitoring of eight state dimensions across five governance areas managing 28 million lines of code. A five-tier action loop (OBSERVE, REQUEST_REVIEW, REINFORCE, SUGGEST_OVERRIDE, ESCALATE) ensures progressive autonomy, while a dual-signal anomaly detector combining Mahalanobis distance with softmax entropy prevents neural overrides on out-of-distribution states. The 128-dimensional feature encoding, LSTM architecture (hidden_dim=512, num_layers=2), and outcome verification feedback loop are described in full. Production deployment achieves 92.96 reliability score with sub-second inference on consumer hardware.

Fields of Study & Subjects

Safety Systems Engineering in Autonomy Software System Performance and Reliability Software Engineering Research

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN