Raw-gyro geographic prediction companion
This is a descriptive comparison using the existing 53 saved selections. It performs no new fit, source decoding or scientific classification. The original raw gyro plot and all earlier findings remain unchanged.
Prediction and observations
Compute latitude/longitude rates from consecutive saved primary-receiver fixes, unwrapping longitude and rejecting unsupported timestamps or gaps over two seconds. Candidate metrics give the north/east velocity and ground course: the existing globe metric includes height; the declared disc uses its fixed radial metric. Keep both courses rather than silently choosing one as measured aircraft heading. Predict the local frame's transport vector and, separately, the specified background field. Upstream removal subtracts the same declared conventional vector for every candidate.
Project onto the aircraft right axis with declared constant roll, pitch and heading-minus-course. Include the corresponding course-turn contribution for constant tilt/crab. Actual independent heading, changing roll/pitch, mount correction and onboard processing remain unknown. Residual motion is not an independently measured aircraft-pitch correction.
Match retained gyro-bin UTC-second centers to unique receiver-interval centers within 0.55 seconds. Preserve bin durations, association error and excluded-bin reasons; no science measurements are interpolated or filled. Read only cached raw aircraft-axis gyro values and durations, never fused attitude/rate corrections. Reject unavailable recorded mounting. Compare signed right-axis rates and their accumulated angles on exactly the same matched support.
Use every selected stretch and preserve source hash, exact interval, record type, mount-configuration identity, coverage and conservative split/shared-leg groups. The 53 curves are not independent replications.
Fixed scenarios
Three candidates; removal fractions 0, 0.5 and 1; constant crab -10/0/+10 degrees; roll -3/0/+3 degrees; pitch -3/0/+3 degrees. That is 243 fixed scenarios per stretch. The level/zero-crab/zero-removal scenario is the transparent reference, not a fit. Displayed ranges cover only this finite scenario grid, not arbitrary aircraft motion or a confidence interval. Do not choose a preferred scenario or estimate a field magnitude from these comparisons.
Qualification and finite limits
Before evaluating preserved observations, pass 48 independent known-answer controls: three candidates, two hemispheres, four cardinal courses and two removal endpoints. Independently differentiate geographic basis vectors and project the physical rotation axis; do not use the prediction implementation to generate its expected answer. Additional unit tests cover tilt/crab, mounting signs, longitude wrapping, time matching, gaps and fused-channel exclusion.
The observation stage admits all 53 saved selections, at most 1,000 cached seconds and 1,500 receiver fixes per stretch, and exactly 243 scenarios each. Reserve identities before execution. Maximum 48 generated controls, 53 observed summaries, 12,879 stretch/scenario evaluations and 12,879,000 bin/scenario projections. One worker/four numerical threads, 180 cooperative seconds, no signals, retries, original reads, native integrations or optimizer starts. Failed controls block the observed stage. Unsupported observations remain reported, never silently omitted or retried.
Freeze inputs, code, scenario identities, numerical environment and reference controls before execution. Publish CSV summaries and a signed observed-versus- predicted plot, with source/group metadata and explicit motion assumptions. Relative scatter is descriptive: report no p-value, independent-vote count, physical-model exclusion or cause of the retained background field.