Structurally Inseparable False Triggers in Movement-Based Early Detection of Physical Hazards in Crowds, and a Posture-Proxy Method for Measuring Them

Abstract

Background. The research prototype for early hazard detection in crowds marks present, observable physical movement; it does not identify persons, infer intent, or predict future offending (working definition, not yet externally confirmed). Problem. Where conduct linked to protected characteristics and a hazardous state produce the same present movement, a marker defined only on that movement cannot, in principle, separate them. We reproduced four couplings on the unmodified prototype's real code path. Prostration satisfies every condition of the fall marker, and a person kneeling beside it satisfies the "assisting" condition, so a group at prayer looks like a medical emergency. Because the prototype's deployment-context gate deliberately exempts life-safety markers, they remain active even in a declared assembly context. Hands raised above the head in prayer satisfy the raised-arm marker. Frequent rests, as with a walking aid, can saturate a loitering indicator. The trap. The intuitive remedy, recognising the prayer posture or the walking aid and suppressing the marker there, would infer religious practice or health and thereby touch GDPR Article 9 and, for religion, potentially AI Act Article 5(1)(g). Such a false trigger should therefore be defused in its consequence; we rule out repairing it by making the system recognise the characteristic behind it. Method. We propose measuring false-trigger rates per body posture, never per characteristic: seven observable classes (torso horizontal, kneeling, wrist above head height, seated, slow gait, stationary in a moving crowd, partial occlusion). How often a bearer of a protected characteristic adopts a posture stays deliberately unmeasured, because measuring it here would require collecting the characteristic. Practitioners could use the same classes to test such systems. Status. Couplings reproduced with synthetic poses and dwell times. No field rates are claimed. A methods contribution, not a results report. Status. This record fixes the text of a conference abstract as submitted on 11 October 2026 to the AETAP 2027 call for abstracts (London, 20 to 23 April 2027). At the time of this record the submission has not been reviewed; acceptance is not claimed. The abstract is a methods contribution: no field rates are claimed. The record contains the text only; the research prototype itself is not published. If AETAP accepts the abstract, a new version of this record will say so. Provenance. The submitted text carries an RFC 3161 time stamp from FreeTSA (serial 0x091BAC63, 11 October 2026, 07:35 UTC) and an OpenTimestamps proof attested in Bitcoin block 970887. SHA-256 of the abstract text as submitted: 01087bf8ceeced4e9e434719d1427fbd38de8a98c32b5e071826c282734ffa6a. SHA-256 of the title: 707e13c0dc9b5ef3f422bc9daa8513a8b43efcf898322966fcfe42159930c6a6.