Python tool for aircraft Doppler audio geolocation using historical ADS-B flight data (Source-available)
Independent four-flyover fixture
This is a supplementary development regression. The primary integration agent's frozen canonical 3/4/6-event renderer owns the acceptance-upload inputs. This fixture does not replace that corpus or request a second national acceptance sweep.
Generate into a new directory:
PYTHONPATH=src .venv/bin/python scripts/make_multievent_benchmark.py --output results/four-events-new
This emits a 520-second MP3/WAV with four independent emission-domain flyovers, 1,812 trajectory rows spanning twelve opaque flight IDs in three distant regions, a normal upload request and a separate truth directory. Seed 502013 is fixed. The first two events have similarly shaped distant decoys; later decoy timings and geometry differ. All events share one recording clock. Harmonics remain one flyover. The inverse search physics is never imported by the renderer.
inputs/request.json contains only audio, the explicit synthetic recording clock
and CONUS coverage. It contains no event windows, bands, chosen flight IDs, receiver
coordinate or search box. inputs/flights.csv is the entire toy flight universe;
build the operator's FlightIndex from all rows without an aircraft allowlist.
Declare it synthetic complete input coverage, not an actual historical receiver
archive. Constant 122 m ground is a synthetic terrain fixture. The true recorder
coordinate and expected associations occur only in truth/labels.json and must
be unavailable to the upload backend. The global search grid remains independent
of the recorder. This is a development case, not a new held-out real recording.
The initial generation is preserved at results/w4-four-event-20261103.
Use results/w4-four-event-v2-20261103 for the current matching fixture: its
canonical altitude_basis=ellipsoidal matches the actual WGS84 renderer. The MP3
bytes are identical to v1; no detector parameters or waveform changed. The local
operator/config.json indexes all twelve trajectories and supplies constant
122 m ground plus a zero geoid in the simulated common datum. These two tiny
rasters are synthetic fixtures, not Earth terrain/geoid data. No receiver
coordinate or geographic search box is supplied to the operator.
The actual W1 discovery adapter ran with both truth directories masked under bubblewrap, using the approved unchanged defaults. Post-run evaluation found 4/4 events, 0 misses, 0 false alarms, with four distinct harmonic groups. Full predictions and labels remain separate. A discovery-only result is inconclusive for location; these counts do not establish association or localization accuracy.
The normal product upload must use all four supported groups, with one shared clock/location. If a configured event or work budget prevents this, record incomplete evidence; do not silently keep a smaller subset. The input and generator hashes are pinned in the generated provenance. MP3 byte identity also depends on the installed encoder; retain each generated hash rather than overwriting a run.