Recording-scale usefulness and numerical error
The short-history component is complete; preserve its frozen sources and results. Next assess full-range bound growth on generated continuous histories at the original 200 Hz rate, with checkpoints up to four minutes.
Use the declared physical candidates and complete broad calibration region. A separate explicitly narrower box may show sensitivity, but cannot replace the original region or be called measured calibration. Retain shared parameters, initial-state uncertainty, timing support and lever-arm observations. Publish unsupported-domain crossings and overly wide bounds as outcomes rather than shrinking regions to obtain a favorable result.
Use independent known-motion references to check native numerical error over long histories, including sampling phase and packet-boundary observations. Record the difference between a mathematical real-arithmetic bound, an empirical numerical discrepancy and an assumed arithmetic allowance. No sampled maximum is a formal whole-program rounding proof. Missing or unsupported allowances must remain explicit in uncertainty outputs.
Freeze a finite generated-input study before execution, with a cooperative wall limit and no signals, archival fits, protected observations or retries. Its result will determine whether the current global-envelope approach is useful at recording scale or needs a separately qualified tighter construction. Shared observational covariance and held-out adequacy remain subsequent gates.