Product · Safety bound generator for ML decision modules

Exact, regulator-ready safety bounds
for machine-learning decision modules.

1.0 · Description

Type-approval of learned components in safety-critical control stacks demands bounds that hold by construction, not by sampling. Zonotropic generates closed-form activation envelopes around any ML decision module, paired with a regulator-ready evidence pack on every run. The envelope is verifiable offline and runs alongside the model in production at sub-frame latency.

1.1 · Reference application

Worked example: Tier 1 Autonomous Emergency Braking activation module. 97-test regression suite — pass-rate baseline available on request.

0 10 20 30 40 50 |u| (m/s²) 0 2.5 5 7.5 upper lower Fig. 1 · activation envelope, AEB module closing velocity (m/s)

1.2 · Evidence pack

Every run produces a regulator-ready submission bundle. Reproducible from artifacts and traceable to source data via content hashes.

1.2.1 Certificate JSONL hashed + signed · input box, output bounds, pass/fail
1.2.2 Parity plots PNG / SVG teacher vs. certified-inference output across the ODD
1.2.3 Operation count CSV per-decision cost and energy proxy vs. baseline
1.2.4 Cover document PDF auto-generated, regulator-formatted submission cover

2.0 · Specifications

2.1 Certification pass rate 100 % 97-test public regression suite
2.2 Inference latency, typ. 0.11 ms commodity x86, single thread
2.3 Inference latency, max. 0.18 ms measured, n=106
2.4 Bound construction exact — closed-form, tight
2.5 IP status pending — US provisional, filed Apr 2026
2.6 Reference application AEB — Autonomous Emergency Braking

3.0 · Roadmap

§ Application Standard Status
3.1 Automotive AEB UNECE R152 / Euro NCAP available
3.2 Industrial cobot velocity limiting ISO 10218 / TS 15066 validated
3.3 Medical robotics FDA 510(k) planned
3.4 Aerospace decision modules DO-178C / DO-254 planned
3.5 Drone collision avoidance — exploratory