Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,128 @@
+---
+license: cc-by-4.0
+task_categories:
+- audio-classification
+- object-detection
+tags:
+- drone-detection
+- counter-uas
+- acoustic-array
+- self-calibrating-threshold
+- cue-to-track
+pretty_name: "nightjar flight 2026-08-13 — first sortie with a self-calibrated presence threshold"
+---
+
+# nightjar flight 2026-08-13 — self-calibrated presence gate, fast cue chain
+
+First field session where the acoustic presence threshold was **measured from the site's own
+drone-free preroll** instead of using the 2026-07-27 constant. Also the first with a fast-reaction
+cue gate. Both arrays recorded at full native depth.
+
+## ★ Headline: the presence threshold was ~8 dB too strict, and it is not static
+
+Calibrated live from 151 drone-free ambient blocks (150 s, band 100–1200 Hz):
+
+```
+TRACK  8.87 dB (p99)     was 14.6      ambient median 3.67, max 10.98
+SEARCH 6.93 dB (p95)     was 10.0      not clamped
+```
+
+The shipped 14.6 dB constant was this site's p99 on **2026-07-27**. Today's site measures
+3.67 dB median — the gate was roughly 8 dB above where the noise actually sits, which is why
+earlier sorties detected on so few windows.
+
+⚠ **And the ambient does not hold still.** `ambient_drift` telemetry (a rolling p95, logged once
+a minute, 20 samples) ranged **7.15 → 15.75 dB**, wandering up to **8.82 dB** from the frozen
+6.93 preroll value. That is far past the ~2 dB that would justify rolling recalibration.
+
+⚠⚠ **BUT do not read that as pure ambient drift.** The drone was airborne for much of this
+session, and `ambient_drift` measures the rolling p95 of the presence statistic *whatever the
+source* — a loud drone inflates it directly. This log therefore **cannot separate ambient
+non-stationarity from the drone's own contribution**. The honest next step is a drone-free
+recording of similar length; until then the 8.82 dB figure is an upper bound on drift, not a
+measurement of it.
+
+## Cue chain
+
+| | |
+|---|---|
+| candidates | 132 |
+| **posted** | **50** |
+| az-masked (site emitters suppressed) | 16 |
+| cue az | median 144.9, sd 44.8 |
+| presence at cue | median 12.0 dB, min 7.7 |
+
+Mid-session the gate was retuned live for reaction speed (daemon restarted into the same
+session; video/blackbox never stopped):
+
+```
+M_AGREE           4 -> 3     agreement window fills a second sooner
+CUE_MIN_PERIOD_S 12 -> 4     corrections every 4 s instead of 12
+PRESENCE_DB pinned 8.87/6.93 (skips the 150 s preroll on restart)
+```
+
+Measured effect: posts began landing at the **4 s floor** (gaps 4.0, 5.0, 4.0 s) versus ~20 s
+before. The pre-change timeline is in the log: first faint contact at t+359 s, first POST at
+t+367 (**8 s**, dominated by the 4-of-5 s agreement gate), second POST at t+381 (the 12 s
+refractory). Longer gaps afterwards are `CUE_SUPPRESS_DEG` correctly staying quiet while the
+camera was already on target — not a failure.
+
+`AZ_MASK="281-297,319-335"` suppressed the utility-pole transformer (array az 289.2) and
+emitter-B (327.5). ⚠ The mask is expressed in ARRAY azimuth while the emitters are fixed in the
+WORLD, so it silently stops working after any registration change — see
+`~/Desktop/TRANSFORMERS.md`.
+
+## Tracker (blackbox, 21.6 min, 3,763 samples)
+
+```
+SEARCH 1750 | TRACK 809 | COAST 365 | ACQUIRE 250 | IDLE 571 | REACQUIRE 18
+track active 1431 samples (38%), of which 937 YOLO-vouched
+pick source while tracking:  blob 520 | yolo 289
+focal while tracking: median 9.2 mm, max 83.1 mm (19x)
+```
+
+⚠ **No GPS truth was captured for this session** (no DJI FlightRecord exported at time of
+writing), so **none of the tracking above is confirmed to be the drone**. On 2026-08-12 an
+identical-looking 38%-tracked session turned out to be on-drone only inside ~25 m and 0%
+on-drone beyond it. Do not cite these numbers as detection performance without scoring them.
+The test that settles it is differential — correlate CHANGES in tracker bearing against CHANGES
+in GPS truth bearing; absolute-error scoring is circular.
+
+## Registration in force
+
+```
+UMA-16      az_offset 302.2   invert_az true    (JBL far-anchored, 42 dB SNR)
+SB-POLARIS  az_offset 339.55  invert_az FALSE   (2 beacons, handedness margin 2.45 vs 50.85)
+```
+
+## Contents
+
+```
+audio/uma16_<epoch>.wav   UMA-16, 16 ch, 48 kHz, S32_LE (32-bit, verified from headers)
+audio/sb_<epoch>.wav      SB-POLARIS, 8 ch, 48 kHz, S32_LE
+audio/cue_log.jsonl       every candidate, mask event, presence_cal and ambient_drift record
+audio/uma.log             daemon log incl. the calibration line
+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, blob_seeded
+blackbox/frames.tar       9,495 substream stills
+```
+
+⚠ **Both arrays are S32_LE.** Reading them as int16 splits every sample, scrambles channel
+identity and destroys the spectrum. The UMA16v2 is natively 32-bit; every wav banked before
+2026-08-09 was silently truncated to the top 16 bits.
+
+**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`.
+
+## Known-good replay reference
+
+Presence detection replayed over two GPS-truthed flights, channel-agreement gate applied:
+
+| threshold | flight A | flight B |
+|---|---|---|
+| 14.6 (old constant) | 3.8% | 16.6% |
+| 6.55 | 67.6% | 66.3% |
+| 4.71 | 84.1% | 79.9% |
+
+These are **window-level presence detections**. A posted cue must additionally clear bearing
+consistency, the az mask and the refractory, so cue availability sits at or below these figures.