Short-term variation in the saved ESA rates
This diagnostic asks how the logged rates vary over short intervals. It uses only the earlier training portion of the already prepared first twenty-minute development interval in recording 187. It does not open original files, later receiver positions or protected flight groups. There are no model fits, trajectory integrations or changes to existing error assumptions.
Preserve all three gyroscope and three accelerometer channels in their logged engineering units. Choose the training cutoff from the frozen query layout. Reject nonfinite values, reversed timestamps, unsupported gaps or a changed preparation/source binding. Retain the actual timestamps; do not interpolate.
At fixed sample strides 1, 3, 30 and 300, form disjoint triples of rate samples. Subtract the straight line between the endpoints from the middle sample, using actual elapsed times. Divide by the Euclidean norm of these three weights. This normalization would preserve the variance of independent, equal-variance point errors. That error law is a hypothetical interpretation, not a property established for the logged rates. Triples are disjoint within each stride; results from different strides are dependent.
Report the sample counts, actual time separations, RMS and median absolute contrasts, robust upper-tail size, means and ranges. Keep angular rates and specific force distinct. No expected rotation field, fused attitude or fitted aircraft motion is subtracted. Aircraft motion, vibration, filtering and noise can all contribute. Linear motion in time cancels; other motion need not.
Do not convert logged point-rate variation into an angle/velocity random-walk specification. In particular, the linear-point integration used for ESA creates shared-endpoint correlations that are absent from independently perturbed synthetic increments. A physical noise law needs a separate sampling/filtering model. This diagnostic supplies no calibrated confidence or instrument limit.
The allowance is one saved-data identity, at most 360,100 input samples and the resulting earlier training prefix, four fixed strides, one numerical thread and a five-minute cooperative ceiling. Charge input samples and each contrast before calculation. No retries, native calls, optimizer starts, original reads or protected reads are allowed.
Qualify independent arithmetic examples, irregular timestamps, linear-trend cancellation, signed axes, disjoint indices, malformed input, preservation of inputs and training-only selection before execution. Bind source and runtime versions. Audit saved outputs and hashes without repeating the numerical work. Any uncertainty-model revision requires its own frozen qualification; neither the active noise test nor its accuracy criteria may be changed.