Measured against what the hardware can do
Edge agents run a DSP chain on RTL-SDR-style I/Q (adaptive noise floor, detection, features, an open-set 'unknown' class) and upload Ed25519-signed batches through a gateway with replay windows, quarantine and backpressure, into SQS, a batched idempotent worker and Postgres. A 51,802-parameter CNN trained on impaired simulated I/Q scored 0.976 accuracy and 0.973 macro-F1 on receivers it never saw in training, and only ships as a signed ONNX model if it beats the baseline's macro-F1 by 0.02.
ADR-0005 makes the live system report localization as unavailable: NTP gives millisecond timing and TDOA needs about 10 ns. The localization engine (RSSI, TDOA, angle of arrival, joint solves, a Kalman tracker) runs only on simulated measurements. A 120-trial Monte Carlo per sensor count puts numbers on the gap: with 12 GPS-disciplined sensors TDOA lands within 6 m median, with NTP clocks it misses by 6.2 km. Results are labelled simulated throughout, because synthetic data overstates accuracy.
Timeline
- Backend, simulator, edge agent, gateway, worker, query API, compose stack
- Signed CNN classifier and localization engine
- Air picture: terrain, probability fields, TDOA hyperboloids
Sources and method
Written from the project's code, documentation and commit history, reviewed in November 2026. Business figures come from me. Clips and screenshots are labelled with where they came from; none show customer data.
spectre/README.mdspectre/docs/ROADMAP.mdspectre/docs/ml.mdspectre/docs/localization.mdspectre/docs/adr/0005-honest-localization.mdspectre/models/gate.json