Qualify prediction ranges over recording lengths
The short-example width gate is complete: all 48 regions and six case covers pass. Never resume this study or reuse its unused allowance.
The four-minute input adapter is complete. The separate 50-second growth screen completes all six cases, measuring 312–520 seconds per bound and 71–72 km unconstrained outer radii. Its smaller approximation remainder does not give a narrower full range than the two available historical whole-change bounds. Preserve those complementary constraints when preparing training-conditioned predictions; neither kind of unconstrained range is useful by itself. The explicit-allowance interface is checked for later training constraints; its caller must supply a justified error budget instead of inheriting the short example's 1 mm allowance.
Qualification must cover the entire integration history, not only the unused prediction period. The earlier representative prediction plan spans 438 seconds in total, with 292 training epochs and 147 later epochs. The four-minute inputs used for compatibility and cost checks cannot by themselves qualify that full horizon. Freeze the intended histories and their complete clock support before the next recording-length study; preserve the existing fixed-clock inputs.
- Inspect the earlier four-minute study and its frozen inputs, outcomes and mathematical assumptions. Preserve its incomplete cases and broad/narrow sensitivity outcomes unchanged.
- Define a recording-length feasibility study for the current shared-calibration method, retaining all candidates and processing assumptions. Rebuild source bindings and derive any training-feasible restrictions from the new study's own training inputs. Short-history certificates cannot be silently relabeled for a longer history. Preserve every physical direction and clock-support requirement.
- Cost the work from measured source visits before launch. A short, bounded duration/runtime screen may precede expensive complete-record calculations, but it cannot qualify recording-length usefulness. Freeze all query times, domains, input hashes, runtime limits and acceptance rules separately. Bound-source and clock reports are query-time dependent. Reusing an earlier integration state at another query time requires separate qualification; streaming value/Jacobian equivalence alone does not establish bound reuse.
- Require complete coverage, supported numerical treatment and useful ranges over the declared recording/prediction horizon. Retaining known truth alone is insufficient; enormous conservative ranges do not pass usefulness.
- Apply the explicit error assumptions. Account for receiver error and accumulated inertial noise, or justify any negligible contribution. Assess a shared error-law family without letting a poorly fitting candidate enlarge its own uncertainty or use future data.
V01 remains in progress until those requirements are met. New withheld prediction fits remain V02 work and have not started. Every new study requires its own finite freeze; no completed budget, protected ESA outcome or aircraft measurement is available for an automatic retry or rollout.