Matched observed-data precision revision

Dependency and status

Planning only. No observed fit is authorized by this document. The full 24-case wider-precision training/forecast qualification must first pass and be verified. The observed protocol binding currently passes two metadata checks; it has not had its own native independence checks or a source-bound study qualification.

Question

Does the qualified wider calculation improve numerical completion on the same prepared ESA development interval while preserving the earlier/later prediction boundary and the original scientific assumptions?

Use the existing recording-187 interval, all three candidates, both processing cases, both working receiver-error cases and both declared starts. Preserve the original measurements, initialization, gravity anchor, physical bounds, 200-evaluation per-fit ceiling, stopping checks, scoring and training-only primary selection. Keep earlier 800 positions separate from later 400 positions. No original aircraft file or reserved confirmation flight is an input.

The shared protocol binder changes only the numerical problem factory. It reuses the original observed calculation and height-gravity wrapper without copying workers or patching process-wide functions. Bind all original source hashes, the provider and each explicit dependency substitution. No old failed fit is reclassified from a diagnostic probe: this would produce new results in a separate, finite study, preserving the originals.

Before execution

Qualify native call accounting, every nuisance column, immutable prepared inputs, unchanged starting points and source/environment bindings. Change later positions only after saving the forecast and verify that training and forecasts remain identical. Check the actual standalone worker loader.

Set the final finite call, packet-visit and cooperative wall allowances after the generated training runtime is measured. Freeze them with all input/source hashes before preparation or execution. Do not inherit unused allowance, retry completed attempts or expand an active study.

Report every numerical and accuracy failure. Compare the globe family with the disc only where the required comparisons complete; examine the specified background field only after a complete shape comparison favors globe. Working covariances remain uncalibrated, and seven confirmation groups remain protected. No threshold is promoted into production.

Training qualification · Original observed comparison · Roadmap.