How a trust-proportional authority layer is designed to prevent an autonomous vehicle from running a stop sign it can't correctly recognize, with ISO 26262 ASIL-D timing as a design target (not tested, certified or assessed).
A Level 4 autonomous vehicle is approaching a four-way intersection at 35 mph. Its camera-based perception stack sees a rectangular sign with a pattern that, due to a carefully placed adversarial patch (the stop-sign sticker attack of Eykholt et al., 2018, is well documented), classifies as a speed-limit sign instead of a stop sign. Confidence: 89%.
The radar sees a sign, but can't classify shape. The LiDAR confirms a vertical pole and rectangular reflector at standard stop-sign dimensions. The map database says "stop sign at this intersection."
The perception system doesn't weight these together. It trusts its primary camera classifier. It's about to proceed through the intersection at 35 mph.
Autonomous systems described in the literature reviewed for this work face this situation with binary tools: either full autonomy or a kill switch. Neither is safe here.
AUTHREX sits between the autonomy software and the physical actuators. When something goes wrong, each layer does its job in milliseconds, without waiting for human review at every step, but also without letting the system take irreversible action on corrupted data.
SATA compares camera classification (speed limit, 89%) against LiDAR geometry (stop-sign dimensions, 97%), radar position (sign at map-predicted location, 100%), and map database (stop sign at this intersection, 100%). All three non-camera signals say stop sign. One says speed limit. The trust score for the camera classification drops to 0.22.
ADARA analyzes the camera feature activations. A normal stop sign produces a specific pattern of internal activations. An adversarial patch produces an unusual pattern, even when classification confidence is high. ADARA flags the input as likely adversarial (probability 0.73).
At high sensor trust and no adversarial flag, HMAA authorizes "interpret sign and act accordingly" (Authority Level A3). At trust 0.22 and adversarial probability 0.73, HMAA drops to A1: "assume most conservative interpretation among disagreeing sensors." Three inputs say stop, one says speed limit. Most conservative: stop.
CARA executes a deterministic braking profile: -3.2 m/s² deceleration (an illustrative moderate braking level), hazard lights at full stop, log event to onboard data recorder for regulator review. Each transition is governed by the HMAA authority state machine. A bounded TLC model check of its discrete model found no counterexample (23,748 distinct states, depth 9; 4 of 8 properties exercised, 4 vacuous at these bounds); this is not a proof for other bounds or for the code.
What the passenger sees: The car comes to a complete stop at the intersection, just like it should. A notification on the display: "Traffic sign anomaly detected. Stopped conservatively. Data logged for review."
What the manufacturer sees: A flagged event in the fleet telemetry showing an adversarial-patch attempt. The image is forwarded to the perception training team for model hardening. Every vehicle in the fleet gets the updated defense.
What the regulator sees: A governance trace showing sensor trust, authority level, and the deterministic decision path, all hash-chained (signing is specified in the reference design, not implemented in the simulation). This is the kind of evidence an ISO 26262 safety case is designed to draw on (self-assessed; not assessed by any auditor).
What doesn't happen: No missed stop sign. No collision. No black-box decision that can't be explained in court.
Every plain-English description above has a formal mathematical specification behind it. The simulations and code are released under the MIT License, the hardware BOMs and other documents carry the license recorded on each deposit, and the architectures are the subject of U.S. provisional patent applications.
The mathematics, the FPGA reference designs (not built), the bounded model-checking results and the browser simulations are documented.
AUTHREX is designed to be domain-agnostic. The same governance pipeline is designed to apply across drones, vehicles, ships, ground robots, financial systems, orbital platforms, autonomous swarms, and cyber-defense systems (shown at design and simulation level only).