Skip to content

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

8.3 KB
Raw Blame

Model-support loss in the tiled search

The source data's spatial/time availability and the model's usable support are separate. The search report retains data_scope_complete for the former and adds model_support_complete for the latter. Any typed unsupported CPA trajectory or unclassified scoring envelope makes model support false. Separately, model_assessment_complete records whether every finite scoring envelope is evaluated, rejected, or covered by a validated score bound. Unsupported CPA trajectories, unreported clock trials, and unbounded alternatives keep the search partial. A completed tile does not establish complete evidence.

For robust-ridge subset association with a bounded recording-start window and no silence filter, an unsupported trajectory that starts strictly after every possible event CPA arrival cannot participate in an association. Nor can one whose latest emission plus the maximum sound-travel delay is strictly before every possible event CPA arrival. Those tracks are reported separately as clock_excluded_unsupported_tracks in stage diagnostics and counted in clock_excluded_unsupported_track_evaluations; their original motion failures and conservative clock bounds remain inspectable. The bound includes the configured clock sweep, CPA and association tolerances, and outward floating-point rounding. Overlap, missing or invalid clock information, fixed-clock mode, other scoring modes, and silence filtering retain the usual unsupported/partial behavior. This does not repair the source trajectory or establish geographic coverage. The HTTP evidence verifier checks the recorded inequality against the search clock and stage totals; it does not independently reload the original aircraft index to verify the recorded emission endpoints.

After these clock exclusions, a tiled CPA stage can prove every unsupported native interval spatially irrelevant to its actual receiver set. Both the native spline hulls and every original CPA sample edge that touches those intervals must pass strict distance bounds. The original spline, sampling lattice, global sample step and identity remain intact. See motion support for the proof and numerical limits. Failed proofs keep the original whole-track unknown envelope. Successful proofs appear as interval_supported_cpa_tracks in stage diagnostics, with their original invalid intervals, receiver box, sound speed, sampling step and distance bounds. The report counts these evaluations in interval_supported_cpa_track_evaluations without counting them as unsupported CPA work. Acoustic trials still check arrivals from those same invalid intervals and remain unknown when their receive envelopes overlap.

The HTTP verifier checks each recorded interval-proof inequality, configuration scope, coarse checkpoint, and aggregate count. It does not reconstruct the source spline or receiver grid from the aircraft index. This proof narrows a CPA loss envelope; it does not establish source correctness or complete spatial coverage.

The additive model_support report contains these integer work counts:

  • unsupported_cpa_track_evaluations: trajectory/stage CPA failures, including repeated evaluations in different coarse/refinement stages.
  • unclassified_score_candidates: candidate envelopes whose work is unknown.
  • evaluated_trials, rejected_trials, unknown_trials: totals explicitly reported by robust scorer assessments.
  • unreported_trial_envelopes: interrupted or missing/malformed trial accounting; its unknown trial count must not be inferred as zero.
  • bounded_score_candidates, unbounded_score_candidates: the partition of unclassified scoring envelopes by whether every scheduled clock trial has a validated score ceiling or a supported-event rejection.

For the normalized subset score, an unknown event contributes at most one unit because its robust loss is nonnegative. A validated complete event can tighten that ceiling; an event exceeding its unchanged residual limit rejects that association at that clock. The maximum bound across every scheduled clock is persisted separately from measured scores. Raw missing-support diagnostics remain unchanged. See the finite-search decision.

The product accepts this accounting only when the raw counts, persisted bound counts, completion flags, and combined discarded-score bound agree. It still defers competitive bounds to the existing confidence decision. No bound repairs missing flight coverage, invalid motion, omitted tiles, event caps, or an incomplete sensitivity profile.

diagnostics_directory points to support/. Per-stage JSON files preserve every affected track ID/reason and every unclassified candidate's receiver, event/flight assignments, pass times, available clock bounds and scoring assessment. Coarse checkpoints also retain these diagnostics; resumed counts replace work rather than inflating it. Refinement failures use separate stage files. Per-stage lists are bounded by the existing track/candidate budgets; the report aggregates counts without collecting a national diagnostic list in memory.

Only geometry.AcousticSupportError is isolated. A CPA error conservatively excludes that entire track's work for the receiver stage, including partial rows from earlier chunks, and keeps valid other-track rows. The diagnostic labels the whole track time span and receiver extent as a conservative loss envelope: this narrow exception need not identify the exact unsupported interval. Optional emission/receive intervals reported by the physical validator are preserved in reported_intervals; they do not narrow the conservative coverage envelope. Original index payloads are retained. Robust scoring classifies typed failures per clock trial; the outer handler also preserves other candidates if a legacy scorer raises the same typed error. An interrupted scorer's unreported trial count stays unknown. Generic programming/configuration errors retain the existing explicit tile-failure behavior.

This handling does not certify continuous total speed or Newton convergence, repair historical missing derivatives or altitude clipping, identify the cause of bad observations, or establish complete real-flight coverage. Issue #64 retains those open gates and the historical 855/858 result. No old outcome is promoted.

The integrated W3 physical guard validates whole-source continuous motion at both actual CPA call sites and time-local motion/convergence in direct acoustic scoring. A real constructed indexed source with zero knot speeds and a >700 m/s interior peak now reaches this outer typed boundary without a monkeypatch. The mixed-source regression preserves the clean candidate identities/scores, retains the original malformed source, and reports partial model coverage. This verified connection does not repair source observations or establish real-data acceptance; W3 documents the numerical support assumptions in MOTION_SUPPORT.md.

Configured research runs

The configured dopplergeist run path also isolates AcousticSupportError at the track and candidate boundaries. A failed CPA track leaves its whole source time span and receiver stage unassessed. Other tracks retain their original CPA rows. A typed scoring failure leaves that candidate's full proposed clock domain unassessed. Generic programming and configuration errors still propagate.

support/ retains the excluded track IDs, reported physical intervals, receiver envelopes, candidate associations, clock domains, and available trial assessments. model-support.json summarizes those stages. Counts include repeated evaluations of the same track or candidate at different resolutions. Missing trial records remain unreported work. They do not become zero unknown trials.

Legacy brightness and ridge scoring do not supply complete clock-trial records. Their finite scores remain ranked hypotheses with core_assessment_complete: false. Unsupported or unclassified work makes model_support_complete and model_assessment_complete false. Original input provenance remains in the report. The isolation does not repair velocities, observations, or physical model support.

The configured report declares result_scope: configured_research_ranking and accepted_location: false. Empty ranking results retain best: null and the support diagnostics. This path does not certify source coverage, eliminate missing competitors, or make an automatic localization decision.