The method
A detector that beeps is easy to build. A detector that stays quiet when there is an explanation is the real problem. All of Omni-Core is organized around that idea.
The Correlator
Mission profile
One question at startup: “Is the building powered? Yes / No / Partially”. In a powered building, a 60 Hz signature is classified as an electrical source. In a building declared unpowered, the same signature becomes a major anomaly — “electrical signature in a building declared unpowered — check”.
Calibration
At startup, and whenever the operator asks for it, the device learns the room’s signature in ~45 s: median and median absolute deviation of every sensor, with a floor derived from its physical resolution. A median is not dragged by a brief episode: a real event during learning does not blind the device to what it is looking for. The floor keeps a stuck sensor — which returns a constant — from disguising itself as a major anomaly. Verdicts are suspended during stabilization — an accepted trade-off.
The device does not guess that it has changed rooms: recalibrating is the operator’s call. In exchange, a passive “reference may be stale” light turns on when at least 60% of the signals have been deviating continuously for 60 s or more; it suspends nothing and changes no verdict. A named room reloads its signature instantly on return.
Detection and correlation
An event is a significant deviation from the learned signature (z-score), never an absolute threshold. Events from all sensors are cross-checked within a sliding window of about 3 seconds. Frequency analysis of the magnetometer isolates the power-grid signature (60 Hz and harmonics).
The three verdicts
| Verdict | Reading | Reaction |
|---|---|---|
| 🟢 Draft | temperature drop + moving air | information |
| 🟠 Electrical source | magnetic fluctuation with 60 Hz, stable temperature | identified, classified |
| 🔴 Unexplained anomaly | magnetic fluctuation without 60 Hz + temperature drop + still air | alert + haptics |
Every verdict carries a confidence score weighted by the size of the deviations; the threshold is adjustable in Expert mode.
When a sensor is missing
A rule separates the senses it requires from the conditions it tests: “the magnetometer is there and sees no 60 Hz” is evidence; “the magnetometer is missing” is ignorance. When a required sense is missing — sensor absent, failed, or silent past its freshness delay — the 🔴 rule cannot fire: it falls back to 🟠 “undetermined — sensor missing”, naming what is missing. Without the thermal sensor a draft is not ruled out, so the device has no right to say “unexplained”. 🟢 and 🟠 rules deprived of a sense simply do not fire.
The rules live in a versioned file, and every verdict carries the version that produced it: when they are tuned after the first field visits, alerts from earlier visits will stay explainable.
A recalibrating sensor goes quiet
Some sensors recalibrate on their own mid-mission: the thermal camera, for one, periodically closes an internal shutter to recalibrate each of its pixels. Its reading then makes a permanent jump — 0.44 to 1.23 °C depending on the direction of the drift, measured on September 16, 2026 — that would mimic exactly the sharp drop the device is looking for. Rule: a recalibrating sensor goes quiet, long enough for the 3-second window to clear its old reference, and counts as a missing sense in the meantime — the degraded verdict above applies. Accepted cost: about 4 seconds without thermal sensing at every shutter cycle.
The self-interference guard
The first suspect is the device itself. It logs its own state — CPU load, SSD activity, Wi-Fi activity, screen, fan — alongside the sensors. The SSD is there for a precise reason: saving the black box coincides with the alert by construction, and writing to the disk emits broadband radio noise that would otherwise be indistinguishable from a real event. Any event coinciding with a spike of internal activity is flagged “probable internal origin” and excluded from the score. Sentinel detections feed the same mechanism: “electronic presence — probable human cause”.
A.R.V.A. v2 — never manufacture pareidolia
v1 fed continuous radio noise into a transcription engine: a machine for hallucinating words, the opposite of the product’s promise. v2 reverses the order:
- Detect — on the acoustic and demodulated-RF channels, look for events (voice-band energy standing out from calibrated noise, formant structure, repeated patterns); voice-activity detector on the acoustic side.
- Correlate — an audio event has no status until cross-checked with the other sensors in the Correlator’s window.
- Transcribe last — only on segments that passed the voice-activity detector, and always labelled: grey = uncorroborated interpretation, green = physically corroborated.
An RF event on a broadcast frequency is flagged “probable source: broadcasting”. Accepted trade-off: fewer words than a classic spirit box — but every displayed word is traceable in the black box.
The black box
Permanent ring buffer. Every alert saves the raw data of all sensors, the audio and the thermal + NoIR videos over ±10 seconds. Every alert is provable or refutable after the mission. AES-256 encrypted logs.
The success test
In a test room, the device must correctly distinguish a provoked draft, a powered-on electrical appliance, and only flag an anomaly if no natural cause is measured. Phase 1 exit criterion: four scenarios classified correctly 9 times out of 10 — a fan, a corded drill, a magnet on a pole passed alone (silence is then the right answer), then the same magnet with a cold source, in still air (the anomaly must then be flagged). Nothing else gets built before that.