joshruby/nightjar-lawnmower1-20260805

Lawnmower1 — drone + lawn-mower confuser, with GPS truth

A drone DID fly during this recording, and a petrol lawn mower was running throughout. That combination is the point: the acoustic cue gate has been tuned against ambient and against quadrotors in quiet conditions, but never against a real target competing with a loud, broadband, ground-level machine sitting inside the rotor band.

⚠ This session was briefly mislabelled as a drone-free negative before the flight record arrived. It is not. The DJI record places the flight entirely inside the recording window: flight 19:17:36–19:22:22 UTC, session 19:17:53–19:28:11 UTC.

Flight envelope — the best profile of the campaign so far

airborne	286 s
max slant range	669 m (furthest yet; previous best 473 m)
max height	64 m
path flown	2536 m
elevation angle	median 11 deg, p90 74 deg

The elevation spread matters. Every sortie on 2026-08-04 after the first sat at 5-6 deg median elevation — the drone on the treeline, at the null of a horizontal array's sin(el) response. This one reaches p90 74 deg, i.e. genuinely overhead passes, where that array is near its maximum elevation sensitivity.

Recorded by the nightjar rig (Bosch MIC-7130 PTZ + Jetson Orin + UMA-16 and SB-POLARIS microphone arrays), same site and hardware as the nightjar-flight-2026080* series.

session duration	~10.2 min
drone	yes — 286 s airborne, GPS truth attached
confuser	lawn mower, nearby, ground level, running throughout
detector mode	day
cue producer	cheap daemon (SUBHUNT=0), posting enabled

★ The finding: a 205 Hz line clears the notch by 5 Hz

cue_candidate   25        chirp_reject   44
threshold in force:  10.0 dB -> 24,   14.6 dB -> 1
cue line Hz:  min 205   median 205   max 779

The deployed site notch is 100–200 Hz — widened to 200 the previous day to kill a fixed site emitter at 141–170 Hz. The mower's dominant line lands at 205 Hz, five hertz above that upper edge, so it passes the notch and then clears the presence gate essentially every time. Twenty-four of the twenty-five candidates were evaluated at the relaxed 10.0 dB search threshold, i.e. the state-dependent gate deployed the day before, which lowers the bar whenever nothing is being tracked.

This is a mechanism-level result, not "the gate is noisy": a specific machine has a specific line at a specific frequency that the notch just misses.

Which cues are the mower and which are the drone is NOT yet separated. With the drone airborne for 286 of the ~610 recorded seconds, some of those 25 candidates may be real. The GPS record makes the separation possible — score each candidate's bearing and timing against the drone's true position — but that work has not been done here. Do not treat the 25 as false alarms without doing it.

Also notable: 44 chirp-rejects against 25 candidates. The chirp guard uses 8–11.5 kHz energy to veto beacon sweeps; a mower is broadband enough to trip it.

Contents

audio/uma16_<epoch>.wav   UMA-16, 16 ch, 48 kHz, S16_LE, 5-min splits
audio/sb_<epoch>.wav      SB-POLARIS, 8 ch, 48 kHz, S32_LE, 5-min splits
audio/cue_log.jsonl       every candidate, with thr_db + nj_state + line_hz per record
video_segments.tar        1080p native + 768x432 substream, H.264, 60 s MPEG-TS
blackbox/state.jsonl      tracker FSM ~3 Hz: state, pose, focal, detection, track
blackbox/frames.tar       1,383 substream stills, f<epoch_ms>.jpg

Channel identity is not guessable. UMA-16 0-based channel 9 is a dead electronic-floor channel — exclude it, run on 15. SB-POLARIS has 3 live capsules of 8; detect them by RMS (live ~−70 dBFS, dead ~−190) rather than hardcoding, as the live set changes between sessions.

⚠ The two arrays use different sample widths — UMA S16_LE, SB S32_LE. Reading both as int16 splits every SB sample in half, scrambles channel identity and destroys its spectrum.

⚠ DJI stamp sparsity: only 6 of 5,607 rows carry a valid CUSTOM.dateTime (the rest are 1970 placeholders). Anchor ONCE off the first valid stamp and derive every row's time from OSD.flyTime — filtering rows on a valid stamp leaves you with 6 fixes instead of ~5,600.

★ Calibration — USE THESE OFFSETS, not the ones the rig was running

The registration in force while recording was stale (az_offset_deg 301.37, carried over from the previous day) because the camera and arrays had been moved beforehand. The correct values were established by a full beacon re-survey later the same evening:

UMA-16      az_offset_deg 297.644   invert_az true
SB-POLARIS  az_offset_deg 108.311   invert_az FALSE   (3 beacons agree to 1.93 deg)

Apply these when deriving bearings from this session's audio. Registration is applied offline to raw multichannel audio, so it is fully recoverable after the fact — the wavs do not care what the rig believed at the time. What is NOT recoverable is where the camera physically pointed: the live cue-driven slews during this recording were aimed with the stale offset and are wrong by roughly 3.7 deg of array drift plus a ~69 deg camera rotation.

Why these offsets apply to this session. The camera/array move happened BEFORE this recording, and the re-survey happened after, with no rig handling in between — only a beacon (the chair speaker) was repositioned, which changes that beacon's surveyed bearing but not the array-to-camera registration. Sessions recorded on 2026-08-05 from ~13:08 onward therefore share this geometry. Earlier sessions that day (e.g. the 11:44 helicopter capture) fall on the other side of the move and should NOT use these numbers.

What the stale registration did and did not spoil: audio, video, blackbox pose, and the cue-log line frequencies are all unaffected and the 205 Hz analysis above stands on its own. Only bearings and live pointing are affected.

For the diagnosis method — why an invert_az flip means "solving against stale truth" rather than a real mounting change, and how three array bearings plus one surveyed camera bearing identify which beacon is which — see the campaign notes.

Companion recordings

Part of a set of drone-free negatives captured the same day and the night before:

• crickets / night ambient — nightjar-night-ambient-20260804
• helicopter pass — flight_20260805_114425 (same rig, daytime)
• an earlier, shorter mower capture, flight_20260805_130826, which ran while the detector was still pinned to night mode in daylight — its optical detections are junk, though its audio is fine. No drone flew during that one.