Skip to content

Python tool for aircraft Doppler audio geolocation using historical ADS-B flight data (Source-available)

3.4 KB
Raw Blame

Hermes reference audit

User-supplied reference: mojomast/hermes-aircraft-doppler-geolocation, pinned on 2026-11-02 at 2c3ed20577bcce29058d73422dd4f85392d04c24. Its LICENSE is MIT, copyright Kyle Durepos. No plugin installation or source-code copy into our implementation was performed. The reference was cloned under an ignored experiment directory for inspection and its own offline tests.

Relevant findings

  • The README explicitly distinguishes its independent reconstruction from the video's unreleased code. It reports real acoustic tests with manual windows/bands, but no real-flight localization validation or scripted acquisition/terrain/grid workflow. Those recordings have no established sample-zero clock, so they do not resolve our independent real-corpus gate.
  • Its synthetic example includes the true observer exactly on its 200 m candidate grid. A first-ranked point with 0 m error therefore does not establish off-grid precision, automatic national recall or uncertainty calibration. Our independent synthetic generator keeps the recorder off the search grid, with truth separate.
  • The documented unknown-clock single-event error is approximately 1.4 km along the track. This is a useful ambiguity case, consistent with requiring independent flyovers and locked relative timing before accepting our default precision target.
  • score_candidates.py preserves emission-time radial velocity and reception delay. It uses local flat-earth coordinates, one global CPA and RMS misfit where lower is better. Our WGS84/gap-split/multiple-approach higher-is-better robust loss is a different score. Thresholds and confidence interpretations must not be copied between them.
  • Its scorer evaluates whichever observed samples overlap track support and only requires five samples. A direct probe with 100 observed constant-frequency samples and support for only five returns zero RMS. This is not a valid complete event association. Our decision requires >=95% support, checked independently of raw spectral rank. No probability or parameter stderr substitutes for this gate.
  • Its synthetic renderer reuses the scorer's local-coordinate helper. It still transports emission phase to reception time, but does not replace independent analytic/gap/Unix-clock tests or the separate scientific replays in this project.

Executed checks

git clone --depth 1 https://github.com/mojomast/hermes-aircraft-doppler-geolocation.git results/references/hermes
cd results/references/hermes
/home/cinna/aircraft-audio-geolocation/.venv/bin/python -m unittest discover -s tests -v

Result: 6 tests passed in 16.654 s, including its synthetic location test and noise rejection. Original source remained clean. The direct scorer probe returned (0.0, -0.0, 110.0) for 5/100 supported samples. Outputs are retained locally as results/references/hermes-test-output.txt and hermes-partial-support.json. These are checks of the reference package, not additional DOPPLERGEIST pipeline or independent real-data successes.