joshruby/nightjar-flight-20260806-1

nightjar flight 2026-08-06, session 1 — a CUE-GATE FAILURE, kept as the negative control

A drone flew and the rig produced essentially nothing. This session is published because it is the clean "before" half of a same-day, same-site, same-flight A/B against nightjar-flight-20260806-2, where a one-line change fixed it.

★ The finding: the chirp guard vetoed the drone

cue_log events:   cue_candidate 14   chirp_reject 18

All 14 candidates fired in the first 62 s and are a fixed site emitter, not the drone: azimuth pinned 137.5–140.6 deg, line frequency pinned 638.7 Hz (9 of 14 identical to 0.1 Hz), presence 18–24 dB. By the peak-Hz discriminator (wide IQR = drone, pinned = site emitter) that is a machine. It is not the 205 Hz mower or the 141–170 Hz emitter notched earlier in the campaign — 638.7 Hz is a new source at bearing ~139 deg.

Then, at the moment the drone launched, 18 consecutive chirp_reject events, one per second, and zero cues for the remainder of the sortie.

The chirp guard vetoes any block whose 8–11.5 kHz energy exceeds CHIRP_FACTOR (3.0) times the trailing median of that band. It exists only to reject the KT-sync pinger beacon sweeps. No pinger was running. On a quiet evening the trailing median sits at the noise floor, so a close broadband source clears 3x trivially — and a nearby quadrotor is broadband. Same signature as the 2026-08-05 mower session (44 rejects vs 25 candidates).

⚠ This is specific to the SubHunt daemon. cue_daemon.py (the cheap gate) runs the presence test FIRST and overrides the chirp reject when the signal is rotor-shaped (chirp_ovr). cue_daemon_subhunt.py vetoes unconditionally, before presence ever runs. This session flew SUBHUNT=1, so enabling the neural bearing path is what exposed the bug.

Consequence for the tracker

session	18.8 min
SEARCH	96.6 %
ACQUIRE	0.2 %
track active	0.2 % (6 samples, one 2.0 s episode)
frames with a detection	1
max focal	15.17 mm (3.53x), median 4.30 mm

search.scan_enabled is false by design — SEARCH holds and waits for an acoustic cue rather than sweeping. With the cue chain dead, the mount simply sat at the last emitter bearing. From outside this is indistinguishable from a jammed mount; the mount was verified healthy the same evening (three commanded slews landed to within 0.02 deg over a 140 deg pan swing).

Drone truth

dji/FlightRecord_2026-08-06_[20-15-21].* — the flight spans the session boundary:

flight airborne	00:15:21 – 00:28:10 UTC
captured here	00:15:21 – 00:20:08 UTC (287 s)
remainder	in nightjar-flight-20260806-2 (00:25:15 – 00:28:10)
max slant	84 m
max height	51 m
elevation	median 17 deg, p90 31 deg
bearing from rig	172 – 264 deg

The ~5 min gap (00:20:08 – 00:25:15) is when recording was stopped to diagnose and patch the chirp guard. The drone stayed airborne across it.

⚠ DJI stamp sparsity: only 7 of 7,694 rows carry a valid CUSTOM.dateTime. Anchor ONCE off the first valid stamp and derive every row from OSD.flyTime; filtering on a valid stamp leaves you with 7 fixes instead of 7,694.

⚠ Timezone: the Jetson clock is Denver (MDT); the operator and DJI wall-stamps are US Eastern. DJI filename [20-15-21] is 00:15:21 UTC the next day.

Configuration in force

cue producer      cue_daemon_subhunt.py  (SUBHUNT=1, ft16, band 250-2600 Hz)
chirp guard       ON   <- the bug
detector.mode     day
yolo_engine       nightjar_v3.pt   (yolo_conf 0.15 / seed 0.45 / imgsz 1280)
cue.blob_seed     true   <- first flight ever with blob seeding enabled
UMA-16            az_offset_deg 297.644   invert_az true
SB-POLARIS        NOT RECORDED - see below

⚠ SB-POLARIS produced no audio this session: Cannot get card index for SBPOLARIS / audio open error: No such device. The card enumerated normally 40 min earlier and dropped off the USB bus before the session. UMA-16 only. Pointing is UMA-driven, so this does not affect the result above.

⚠ Blob seeding was enabled for the first time here and is NOT validated by this session — with the cue gate dead it was never given a cue window to seed inside.

Contents

audio/uma16_<epoch>.wav   UMA-16, 16 ch, 48 kHz, S16_LE, 5-min splits
audio/cue_log.jsonl       every candidate and reject
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       390 substream stills
dji/                      DJI FlightRecord (txt + decoded csv + geojson)

Channel identity is not guessable. UMA-16 0-based channel 9 is a dead electronic-floor channel — exclude it, run on 15. Load the serpentine geometry from uma16_geom.json; hand- deriving it has faked a 50 deg error before.

Companion

• nightjar-flight-20260806-2 — the same evening with the guard disabled: airborne cues went 0 → 8, so the guard finding here stands. ⚠ But GPS scoring of that session shows its optical "locks" were clutter (placebo-level bearing residual) and most cheap-gate cues were a second site emitter at az ~323 — see that card before citing either session as a working cue-to-lock chain.