ESA prediction result: metres versus kilometres

Three complete comparisons favor globe geometry and the specified background field. In this development recording, that model predicts unused positions with a typical three-dimensional error of about 15 metres. Zero-field globe gives 3.8–4.4 kilometres, and the tested disc 7.4–8.6 kilometres.

These are comparisons of RMS position errors, not calibrated probabilities. All three settings reuse one twenty-minute recording, rather than three independent flights.

What was tested

The recording comes from the ESA airborne survey archive. It contains about 360,000 IMU samples, averages 292.2 km/h and lies near 64.19° N. An IMU (inertial measurement unit) records turning with gyroscopes and specific force with accelerometers.

Earlier receiver positions determine the initial state and instrument calibration. The software then predicts later positions continuously from the IMU record, without using those later positions to adjust the answer. Horizontal and height errors must both meet the fixed accuracy limits.

The four settings combine two assumptions about upstream background-field removal with two assumptions about how receiver errors persist in time. Each setting requires qualified forecasts from all three candidates. Repeated starting points probe numerical stability; they are not additional flights.

What improved

Wider computer arithmetic improves numerical completion from 8 to 22 of the 24 starts, with all eight earlier passes retained. It changes no measurement, calibration bound, starting point, stopping check or accuracy limit. Independent generated controls qualified the revised calculation before this run.

The preferred model's later horizontal RMS is 6.8–7.8 m and height RMS 13.4–13.7 m. It passes the physical accuracy checks in all three complete comparisons. The alternatives' much larger errors remain after their numerical fits have converged.

Two disc starts in the short-memory, profiled-removal setting reached the eight-hour cooperative ceiling. That setting remains incomplete. The run used 1,175 native calculations; its saved-results audit repeats no fit or propagation. Every failure and all charges are preserved.

What remains

This is development evidence from one recording. The working error assumptions still need assessment, and statistical confidence has not been calibrated. Seven separate ESA confirmation groups remain protected. The supplied field rate does not establish its cause or a solar-versus-sidereal distinction.

The next check examines the saved errors' growth and temporal dependence before broader prediction tests. The 53-stretch findings remain unchanged.

Full result tables · Every start · Matched baseline changes · Execution scope · Next analyses.