Skip to content

FGO external Doppler dead-reckoning witness audit

Purpose

Determine whether a Doppler-only dead-reckoning (DR) track can safely promote otherwise-FLOAT LAMBDA position candidates. This is an offline, monitor-only audit. It does not add single-difference Doppler factors to the FGO graph, change an ambiguity decision, seed a hold, or alter a reported position.

The originally proposed current-epoch SPP/LAMBDA agreement audit was not repeated. The completed multi-epoch audit already included a fresh SPP witness within 5 m in its best zero-wrong development rule. That combination recovered only six correct epochs over all 36,346 Tokyo epochs (0.0165 percentage points) and was rejected.

Research and implementation basis

The repository already contains a dormant external Doppler-DR validator. It solves receiver-clock-drift-free single-difference Doppler velocity by weighted least squares, removes rows beyond four sigma once, propagates a 3-D position and covariance, and compares an integer candidate through a 3-D Mahalanobis statistic. Previously this was evaluated only on relaxed-ratio activation paths. --external-dr-shadow extends the telemetry to every provisional LAMBDA candidate while leaving every decision path disabled.

The DR track resets only from a normally accepted candidate with ratio at least 20 in shadow mode. The reset epoch itself is not evaluated; candidate testing begins after at least one Doppler propagation epoch. The frozen availability limit is 30 epochs and the frozen acceptance threshold is chi-square 11.345 (99% for three degrees of freedom).

Frozen development gate

Tokyo run1 is the only development run. A runtime rescue experiment and run2/run3 inspection are permitted only if the fixed shadow rule produces:

  • at least 100 evaluated FLOAT candidates;
  • at least 60 accepted correct FLOAT candidates (about +0.5 percentage points over 11,905 run1 epochs); and
  • zero accepted wrong FLOAT candidates.

A candidate is correct only when its own ECEF position, transformed into the CSV local ENU frame, is within 0.5 m in 3-D. The reported FLOAT position is not substituted for the candidate. Reference truth is read only by the offline scorer after the solve.

If run1 fails any requirement, the experiment stops: do not inspect run2 or run3, do not enable the existing runtime validator, and retain the shadow only as diagnostic evidence. If it passes, freeze every threshold and require zero accepted wrong candidates independently on run2 and run3 before any default-off runtime A/B.

Reproduction

Append --external-dr-shadow to the documented shipping FGO preset and dump the epoch CSV. The monitor deliberately does not require --sd-doppler: single-difference Doppler rows are materialized for the private DR estimator but are not inserted into the graph.

python scripts/analysis/analyze_fgo_external_dr_witness.py `
  --epoch-csv build-ffrt-msvc/validation/tokyo1_external_dr_shadow.csv `
  --json build-ffrt-msvc/validation/tokyo1_external_dr_shadow.json `
  --markdown build-ffrt-msvc/validation/tokyo1_external_dr_shadow.md

The scorer exits successfully only when the frozen development gate passes.

Tokyo run1 result

The 50-epoch authority-neutrality replay matched all status, ECEF position, ratio, fixed-count, and AR-outcome fields exactly. Only the nine intended external-DR diagnostic columns differed.

The full 11,905-epoch run produced 8,216 candidate epochs. The independent track evaluated 252 otherwise-FLOAT candidates:

Population Count
Correct / wrong evaluated FLOAT candidates 115 / 137
Accepted correct FLOAT candidates 112
Rejected correct FLOAT candidates 3
Accepted wrong FLOAT candidates 112
Rejected wrong FLOAT candidates 25

Verdict: FAIL. The support and correct-yield requirements passed, but the zero-wrong requirement failed by 112 candidates. Doppler-only DR followed the same wrong position basin too often to be an independent ambiguity witness on this run. Per the frozen protocol, run2/run3 were not inspected, the runtime validator was not enabled, and no FIX-rate gain is claimed. The default-off shadow remains useful negative evidence and diagnostic telemetry.