joshruby/nightjar-flight-20260806-2

nightjar flight 2026-08-06, session 2 — the fix, and a producer swap mid-session

The "after" half of a same-evening, same-site A/B against nightjar-flight-20260806-1, where the chirp guard vetoed the drone outright. Here the guard is disabled and the gate cues while the drone is airborne (0 → 8). ⚠ But GPS scoring (added after first publication) shows the optical "locks" were NOT the drone and most cheap-gate cues were a second site emitter — see the scored findings below. An earlier revision of this card claimed "the full chain fires"; that claim is RETRACTED.

⚠ This session contains TWO different cue producers. Treat it as two segments, not one:

segment	UTC	producer	chirp guard
A	00:25:15 – 00:42:28	cue_daemon_subhunt.py (ft16 neural)	off
B	00:42:28 – 00:53:27	cue_daemon.py (cheap gate)	off

The swap was made live because SubHunt was starving the cue rate (below).

★ Finding 1 — the chirp guard was real, but it was not the only problem

With CHIRP_GUARD=0 the gate produces 8 cues coincident with an airborne drone (session 1: zero). That part of the fix stands.

⚠ Scored against the DJI GPS record, most of those cues are NOT the drone either:

cue disposition (14 total)	n
fired while the drone was not airborne	6
pinned at array az 322.6–323.4 — a second fixed site emitter	6
consistent with the drone	2 (SubHunt az 184.0 vs GPS 193.0 = −9.0°; one cheap az 164.4 vs 208.9 = −44.5°, marginal)

Best-fit constant-offset residual over the 8 airborne-coincident cues: 56.5° robust sigma, 12% within 10° — the beamformer mostly locked the loudest static source, not the drone.

⚠ Lesson recorded: wandering line_hz is NOT a drone signature by itself. The az-323 emitter cluster wanders 363–1166 Hz (the peak-picker hops harmonics on a stationary machine) while its bearing stays pinned to 0.8°. Bearing spread is the discriminator; frequency spread fooled the first read of this data.

★ Finding 2 — SubHunt starves the agreement window on this hardware

	segment A (SubHunt)	segment B (cheap)
duration	17.2 min	11.0 min
cue candidates	1	13

The gate needs 4 agreeing bearings inside a 5 s window at ~1 Hz. SubHunt logged analysis behind, skipped ~81 hops over 825 hops (10 %). That rate is the documented steady-state, not the retracted "runaway" — but losing one hop in ten still prevents the 4-of-5 s window from filling. The 2026-08-05 preallocated-ring fix is real and present (1 startup overrun, skips flat); it removed the ALSA overruns but did not make ft16 fast enough to sustain the gate under a live tracker.

⚠ Rate is not accuracy. GPS scoring reverses the quality read: SubHunt's single cue was the only clearly drone-consistent bearing of the session (−9.0° vs truth), while the cheap gate's 13 cues were dominated by the az-323 emitter and not-airborne false alarms. n=1 proves nothing about ft16 either — but do not cite this session as "cheap gate beats SubHunt"; on bearing truth it was the other way around.

Tracker outcome

	session 1 (guard ON)	session 2 (guard OFF)
track active	0.2 %	5.7 %
frames with detection	1	83
track episodes	1 (2.0 s)	22 (longest 7.6 s)
max focal	15.17 mm	11.59 mm (2.70x)
SEARCH	96.6 %	89.7 %

⚠ The FSM numbers above are NOT drone locks. Scored against GPS (256 tracked samples with truth within 2 s): camera-implied bearing vs drone bearing has 27.5° robust sigma after removing the best-fit constant offset — a time-shuffled placebo scores 27.2°, i.e. the real pairing beats exactly 50 % of random pairings. The 22 episodes were clutter (treeline / cloud edges), indistinguishable from real locks in the FSM but not pointed anywhere near the drone. An earlier revision of this card counted them as drone acquisitions; that is RETRACTED.

The honest statement: the guard fix restored cueing; nothing in this session demonstrates a verified optical lock on the drone. (Caveat on the scoring itself: rig position was taken from the 07-27 survey and the rig moved 08-05 — at 24–83 m slant a ~10 m rig-position error is up to ~13° of bearing, which inflates the residuals but cannot produce a placebo-level track result or the pinned az-323 cue cluster.)

⚠ cue.blob_seed: true was enabled for the first time on 2026-08-06 and is present in both sessions. It is NOT yet validated — no analysis here separates blob-seeded acquisitions from YOLO-seeded ones. Do not credit it for the improvement above; the controlled variable between the two sessions is the chirp guard, not blob seeding.

Drone truth — two flights

dji/ holds both records. Flight A spans the session boundary; flight B is wholly inside.

	flight A [20-15-21]	flight B [20-37-43]
airborne	00:15:21 – 00:28:10 UTC	00:37:43 – 00:50:47 UTC
captured here	00:25:15 – 00:28:10 (175 s)	784 s (all)
also in	session 1 (287 s)	—
max slant	84 m	138 m
max height	51 m	47 m
elevation	median 17, p90 31 deg	median 26, p90 44 deg
bearing from rig	172 – 264 deg	166 – 302 deg

Operator notes the drone flew approximately the same positions in both sorties, which is what makes the session-1 / session-2 comparison fair.

⚠ DJI stamp sparsity: 5 of 7,846 rows (flight B) and 7 of 7,694 (flight A) carry a valid CUSTOM.dateTime. Anchor ONCE off the first valid stamp and derive every row from OSD.flyTime.

⚠ Timezone: Jetson clock is Denver (MDT); DJI wall-stamps are US Eastern. Filename [20-37-43] is 00:37:43 UTC the next day.

Configuration in force

cue producer      SubHunt (ft16) then cheap gate - SEE SEGMENT TABLE ABOVE
chirp guard       OFF  (CHIRP_GUARD=0; no KT pinger was running)
detector.mode     day
yolo_engine       nightjar_v3.pt   (yolo_conf 0.15 / seed 0.45 / imgsz 1280)
cue.blob_seed     true   (unvalidated)
search.scan_enabled  false  - SEARCH holds for a cue, it does not sweep
UMA-16            az_offset_deg 297.644   invert_az true
SB-POLARIS        NOT RECORDED

⚠ SB-POLARIS produced no audio: the card dropped off the USB bus before session 1 and did not return. UMA-16 only for both sessions of 2026-08-06.

Contents

audio/uma16_<epoch>.wav   UMA-16, 16 ch, 48 kHz, S16_LE, 5-min splits
audio/cue_log.jsonl       every candidate and reject, both producers
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       1,760 substream stills
dji/                      both DJI FlightRecords (txt + decoded csv + geojson)

Channel identity is not guessable. UMA-16 0-based channel 9 is dead — exclude it, run on 15. Load the serpentine geometry from uma16_geom.json.

Open questions this session raises

1. Tune the chirp guard rather than disabling it. Reject records now log hi and hi_med, so an absolute floor (or the cheap gate presence-override) can be fitted instead of CHIRP_FACTOR alone. Porting chirp_ovr into the SubHunt daemon is the obvious fix.
2. The az-323 emitter needs suppression (survey it; a bearing mask now exists in the cheap daemon: AZ_MASK="315-331", logging ev:az_masked).
3. A verified drone lock is still outstanding — next session runs a truth-scored look-away drill (hovering drone at a surveyed station) instead of free flight, so every cue and every lock is scored against a fixed truth pose at the moment it happens.
4. Fresh rig GPS fix needed — all bearing scoring above leans on the 07-27 rig position.

Companion

• nightjar-flight-20260806-1 — same evening, guard ON, the negative control.