Warning before the point of no return: an onboard early-warning architecture for driver-operated high-speed railways

Railway safety depends on recognising a developing hazard while a safe stop is still possible. This paper defines RAILGUARD-AI as an onboard early-warning system whose purpose is to warn the driver before the train passes the point of no return: the last instant at which a response begun after the warning can still stop the train short of the hazard, given the response delay, dependable deceleration and a safety margin. The novelty is not a new sensor or safety principle but the combination of three elements: the point of no return as the system’s defining requirement and evaluation endpoint; onboard cross-checking of movement authority, occupancy, voice-instruction candidates, sensing and protection availability during degraded operation; and a time-indexed register of lost protective coverage. Established automatic train protection retains primacy, and artificial intelligence cannot grant authority, relax restrictions, suppress protection or release brakes. Japanese protection radio, automatic train stop acknowledgement limitations and the 2017 Nozomi bogie incident motivate requirements for warning delivery and communication closure. In a hypothetical stopped-train scenario at 200 km/h, information arriving 6 km ahead yields a warning 62.5 s before the point of no return, whereas forward sensing at 614 m detects the hazard after it. A prospective evaluation protocol with a matched deterministic baseline is specified. No prototype performance, certification or accident-prevention effect is claimed.