Safety Nets Shrink Certifiable Aviation Neural Networks by Three Orders of Magnitude With 100% Correct Runtime Behavior

arXiv 2608.20053·medium signalXBlueskyLinkedInCopy link

EASA requires a Safety-by-Design approach for safety-critical AI in aviation, achievable via Safety Nets that pair a compressed neural network with a lookup table covering its errors to guarantee 100% correct runtime behavior across a discretized operational design domain. This is the first systematic analysis of the network-size versus table-size trade-off, and it lands on concrete design parameters: 3 to 5 hidden layers of roughly 50 to 100 nodes each with one-hot encoding, where the network accurately represents at least 97% of the data and a compact table handles the rest. The resulting Safety Nets cut system size by almost three orders of magnitude, fitting current avionics memory budgets, and ship as the first open-source Safety Net implementation for HCAS and VCAS.