loom-oi: A Hardware-Ready, Innate-Reflex Threat Detector Built on a Simulated Organoid's Own Response Magnitude

We report loom-oi: a simulation in which an organoid's own firing intensity, driven by a single continuous physical quantity, is used by a reward-trained readout to reliably distinguish a calm scene from an approaching threat. The driving signal, frame-to-frame visual expansion rate, is the same continuous quantity that triggers innate, unlearned defensive responses across virtually every visually-capable species tested, via one of the most evolutionarily conserved circuits in the vertebrate brain. Calibrating this system in simulation surfaced a genuine, non-obvious failure mode: a naive control condition's own measurement noise can exceed a real, weak version of the signal being detected. We diagnosed this down to per-class accuracy and corrected it at its source. After correction, a systematic screen found five of fifteen independent random initializations passing a three-stage qualification protocol, and, under an equal real-time budget, the organoid readout reached 86% accuracy against a standard gradient-based digital model's 100%. This paper reports the result of the real-hardware calibration step the original design identified as the necessary next test. Across a four-hour session on a live human neurosphere culture, five independently designed stimulation strategies were tested, and none produced a usable calm/threat signal. A direct comparison between single-electrode and full 32-electrode encoding, at matched repetition count, found the inverse of the relationship the simulation design assumes: a single electrode produced a higher mean response than the full array.