EdgeGate

EdgeGate measures a new perception skill on each edge chip in a home before it ships, and refuses to deploy what misses the rules. The skill, “tell me when a package is left at the entrance”, was taught from our own recordings (SAM 3.1 labels, a distilled YOLOv8n student), compiled for a Jetson Orin Nano and a Raspberry Pi 5 with a Hailo-8, and measured on the device itself over 2 h 50 min of frozen eval recordings whose hash was written into the policy before any candidate was measured. Our best model caught 25 of 26 arrivals at p95 19.5 ms on the Jetson Orin Nano and 15.9 ms on the Hailo-8, but raised 4.24 false alerts an hour on the Jetson (4.59 on the Hailo-8) against a limit of 1, so the gate held it. It also held all twelve builds we broke on purpose, including a 4-bit build that a false-alert-only gate would ship (0.35 false alerts an hour, 8 of 26 arrivals caught). From the same records, always-deploy ships 20 builds, a false-alert-only gate 4, the rules 0. Muse Glimmer, running locally, audits the evidence through MCP tools and can only add a HOLD, never remove one. Alert delivery is measured on one clock to the push server (1.2 s) and to the phone’s flash (1.7–2.4 s). Every record, audit and verdict is in the repo and replays without hardware (CI does it), and the eval sessions are public with people blurred. The problem it solves: a skill that demos well is not a skill that can be trusted on a 3-watt chip at night; the gate makes “ship it” a measured decision with written-down reasons.