UNCLASSIFIED
UNCLASSIFIED

SHIELD/ATLAS — Acoustic capture lab

Browser microphone → WebSocket → FFT classifier. Tag rule: GREEN = best score ≥ 0.55, AMBER = 0.30–0.55 (verify), RED = < 0.30 (decline, called "background/unknown").

Capture controls

Test session

No session bound — classifications will not be scored.
Open this page over HTTPS (or localhost) — browsers refuse getUserMedia on plain HTTP. For static-platform runs, hold the mic at a fixed distance, switch ground truth to match the playing source, and let the session collect 30+ observations per class for a useful confusion matrix.

Live classification feed

0
Windows
0
Matches (vs GT)
Live accuracy
No classifications yet. Start the microphone to begin.
Click after finalizing the session.

Replay a recorded WAV

Replays a local WAV file (16-bit PCM, mono or stereo, any sample rate) through the same classifier the live mic uses. Use this to score the windowsill recording you captured separately as backup evidence. The path must resolve under the project directory — drop the file into docs/intel/test-corpus/ first.

Corpus calibration

Walks a directory of labeled WAVs and runs every file through the classifier into the bound test session. Filenames or subdirectories matching known platform IDs (DJI-MAVIC-3, SHAHED-136, etc.) are auto-labeled; everything else falls back to "background" unless overridden. Generates a real-corpus precision/recall metric independent of the synthetic self-test.
CHEATSHEET
ASK AI
PRESENCE
⚠ SANDBOX / TRAINING MODE — Live read-only data. Write commands are inhibited (train as you fight, missile button safed).