joshruby/oren-uma-denver-202512

Oren UMA Flights — Denver, 2025-12-31

The earliest era of the acoustic drone-detection estate ("acoustic dragon" / AcousticTriangulation period): two UMA-16 recording sessions with DJI truth, flown near Denver, CO (~39.6256 N, −104.8523 W) on 2025-12-31. Recovered 2026-08-30 from the maxone external drive (Mac_Offload_2026-06-28/) and iCloud; these files existed nowhere else — not on HF, not on any compute box.

Contents

file	what it is
Oren1_complete.tar (68 MB)	Flight 1 truth + metadata: Oren1.DAT (DJI raw), FlightRecord, array_calibration.json, alignment_full.json, config/recDefs/log, timing diagnostic, Airdata CSV
Number1.wav (1.6 GB)	Flight 1 UMA multichannel audio (recovered from iCloud; the only copy that existed)
Oren2_complete.tar (99 MB)	Flight 2 truth + metadata (same structure + KML). Flight 2 audio has NOT been located — searched Mac, iCloud placeholders, and both external volumes; if a Number2.wav-like file surfaces, it belongs here
oren1_reference_library.tar (45 KB)	Range-binned clip-library SKELETON (clips_5m…clips_50m + metadata.json) — structure only, no audio inside
CSVDJI1.csv / CSVDJI2.csv	DatCon-decoded DJI DAT logs for flights 1/2 (fine-grained truth)
DatCon.zip	the decoder tool used, for reproducibility
GPS_PPS_timing.txt	timing notes from the era

Notes

• These predate the Pennsylvania rig: different site, different mount, its own array_calibration.json. Any use for eval/pretraining needs the standard frame audit (fit one rotation per session, positive classical control) before labels are trusted — see eval/audit_ledger/ conventions in acoustic-knowledge.

• Era context: the "Oren-validated" Mamba2 DOA weights (Mamba2 repo) trace to this data.

• CAVEAT: the CSVDJI decodes contain mis-parsed frames (garbage float rows with raw hex) — treat Airdata CSV + raw .DAT + DatCon re-decode as the trustworthy truth chain.

• RTK: YES — Mavic 4T with RTK (correcting an earlier note: Airdata's template just omits RTK fields). Verified empirically from the log: 30-second hover windows show 0.9–2 cm median / 2–4 cm p95 position scatter — RTK-fixed, not consumer GNSS.

• Truth budget (measured, from the era's own receipts): drone position 2–4 cm (negligible in degrees); audio↔GPS time alignment 5.9 µs RMS over 1,044 markers (alignment_full.json); array frame calibrated to heading 90.000° with 1.516° RMS over 10 measurements (array_calibration.json). Net: a ~1.5°-class bearing referee out to 235 m — the highest-accuracy truth chain in the estate, and close to the metrology-grade referee that (per the 2026-08-30 deep-research sweep) does not exist in any public corpus.

• Pending before ruler use: the standard frame audit (classical positive control + one fitted rotation, expected ≈ the calibrated 90.0° heading) to confirm the chain end-to-end on the audio itself.

CERTIFICATION RESULT (2026-08-30, self-localization audit)

PASS-NEAR — referee-grade for 0–50 m only. Full receipts: eval/audit_ledger/oren-uma-denver-202512__oren1.json + eval/oren_scores/summary.json in acoustic-knowledge; the eval cache is baked from the ACOUSTICALLY SOLVED pose, not the calibration file.

• Corrections to this README: the wav is 17.4 min (1,044 s), S16 (not ~27 min); ch9 is healthy in this era.
• The self-loc solve (dE/dN/τ/rotation vs RTK, holdout-validated) REJECTED the era calibration: origin off by 17.3 m (the era's own manifest implies a different origin too — the calibration coordinates were never actually used), heading off ~7°, truth clock offset +0.714±0.235 s. Solved pose halved the 0–50 m error (17.5°→8.95°, tracking subset 5.91°).
• Beyond 50 m the audio is dominated by two stationary broadband confusers (world az ~25°/280°, coherence ~0.8); no frequency band rescues 150 m+. The site, not the truth chain, sets the range limit.
• The reflective-table concern was measured and cleared for bearing (no coherence comb, no per-band azimuth dropout at 0–50 m). Elevation was never scored. Era-manifest elevations are garbage (implied array 18 m above ground) — use only its (t, lat, lon) as a cross-check.

---

Findings and usage rules (2026-09-06 campaign audit) — READ BEFORE USING THIS DATA

Truth and alignment (verified)

• Array pose to use = acoustically SOLVED pose: era array_calibration.json coordinates (39.62588885, −104.8523753) shifted by +0.22 m E, −17.26 m N (±0.94 m), table height 0.762 m. The era coordinates alone are 17.3 m off (self-loc audit 2026-08-30). Verified 2026-09-06: truth rebuilt this way matches the certified flightbank (joshruby/wsu24-stft-cache/flightbank/oren-uma-denver-202512__oren1.npz, az_world/slant) to 0.0 ± 0.2° azimuth and 0.0 ± 0.5 m slant; the other sign of the offset misses by 9° / 28 m.
• Audio↔GPS time: alignment_full.json — GPS time of sample n = 1767213565.297585 + n/48000 (GPS-PPS electromagnet markers, RMS residual 6 µs). Audio spans 2025-12-31 20:39:25.30 → 20:56:49.35 UTC (1044 s, 16 ch, 48 kHz, Number1.wav).
• Flight truth = Oren1.csv (Airdata): datetime(utc) has 1 s resolution — use time(millisecond) from the first row for timing; heights in feet (height_above_takeoff(feet) × 0.3048); positions 10 Hz. DJI track 20:39:30 → 20:56:31 UTC; 960 s aloft, slant median 60 m, p90 214 m, max 295 m; height max 122 m. Drone-free audio: 21 s before take-off, 38 s after landing only — no usable null in this recording.
• Channel order = the July-2026 UMA-16 convention (rows [8,7,10,9][6,5,12,11][4,3,14,13][2,1,16,15], channel = MIC label − 1, 42 mm pitch, channel 9 (0-based) is the hot electronic channel). Verified on this recording by the pairwise max-|TDOA| distance-matrix test: corr 0.975 with the July geometry, 8.9 σ over random permutations (home reference 0.658, 7 σ). Do not re-solve the mapping.
• Flight 2 audio is still missing (Oren2_complete.tar has truth only).

How good a bearing yardstick this is (math)

Angular truth floor = atan(e/R), e = combined position error (array solve ±0.9 m ⊕ drone GPS ±2.5–5 m):

range	floor for repeatability
50 m	~3–6°
100 m	~1.5–3°
150–300 m	~0.5–1° (best in the estate)
The solved heading (~2–3°) is a constant offset: it caps ABSOLUTE accuracy, not precision. Geometry caveat: the far leg (150–295 m) sits on ONE radial (truth azimuth spread 0.6° over 13 ten-second windows) → it measures repeatability only. The inner 150 m sweeps ~145° of azimuth → that is the tracking test. Elevation reaches 55–60° near the array.

What the audio contains besides the drone (this is the point of the dataset)

• The 2026-08-30 audit found two stationary broadband sources with coherence 0.8 in the 50–150 m cell. On 2026-09-06 the validated cube bearing recipe (cube_eval.py --fuse 1 ... --geom uma16_geom.json), which gives 2.2° on a home hover and 12° on a home flight with this same geometry, returned ~90° median here under every yaw offset and under a mirrored layout. A strongest-source bearing points at the site sources, not the drone.
• Usage rule (Josh, 2026-09-06): we do not need the drone to be the loudest source. Track everything, name later — the site sources are simply other tracks. Any bearing method evaluated here must be source-selective / multi-track (steer on the drone's blade-pass harmonic bins, or run K≥3 tracks and score the one associated with the drone's comb), and must be evaluated per track, not as a single strongest-peak estimate. The old "captured beyond 50 m" verdict is a property of single-peak methods, not of the data.
• Planar-array note: conjugating the steering term is exactly a 180° azimuth rotation (not a bug); per-day mirror-sign flips at home are mount handedness. Fit sign and yaw on alternate windows, test on the others.

Presence (detection) use

Aloft/absent labels need only the PPS alignment and 2 m height; range bands 0–50/50–150/150–600 m tolerate the pose error. Used as the held-out-SITE test for the per-microphone detector (trained on Pennsylvania days; receipt pool_v2u/site_eval.json). Because there is no same-site null, that test reports AUC vs Pennsylvania drone-free recordings and recall at a Pennsylvania-calibrated threshold — never a Denver false-alarm rate.

Derived artefacts (private HF dataset joshruby/stf-campaign-202609 unless noted)

• 1 Hz per-mic cubes 60–1600 Hz: finef/oren-uma-denver-20251231__flight1/uma16_1767213864.npz, _1767214164, _1767214463 (also on Spark ~/exp51/finef/).
• Per-second truth (solved pose): truth/windows_long/oren-uma-denver-20251231__flight1.parquet (t, az_tru, slant, el).
• 30 s cube_eval segments with DOATruePos.csv in the solved frame (X north, Y east, Z up): Spark ~/exp51/oren/segs/{fit,test}/seg_NN/ (33 segments; even = fit, odd = test).
• Scripts: code_lab/te_oren.py (cube + truth), te_denver_bearing.py (comb-selective steering + Bayesian state), denver_segs.py, oren_bank_check.py (truth vs certified bank), uma_convention_check.py (channel order), te_site.py (held-out-site presence).
• Ledger rows: receipts_te/EXPERIMENT_LOG.md → "DATA v2", "DENVER far bearing", "DENVER yardstick, settled".