Fixed-point diagnosis of the three horizontal repeat failures
The saved-only audit finds three failed IMU21 repeat pairs despite passing individual fits. The two zero-field globe cases have timing offsets separated by about 19 ms. The unsubtracted flat pair has matching timing and costs but misses the 1 mm prediction-RMS limit. Both linear correction formulations are feasible and predict negligible further improvement at the globe endpoints.
Fixed points and controls
Use both endpoints of all three failed pairs, with no interval/model selection changes. For each of these six endpoints evaluate exactly:
- The saved center.
- The center with the single largest normalized primary/repeat coordinate difference replaced by its counterpart's value. This is timing in the two globe cases and gravity correction in the flat case; freeze this choice.
- The midpoint of the complete primary/repeat parameter vectors.
- The saved range-scaled full affine correction from the completed audit.
Two control identities are the IMU8 unsubtracted specified-field globe primary and repeat. Each receives a center reproduction and its saved affine correction. Every point retains the full parameter vector and original physical bounds.
Frozen allowance and interpretation
- Eight identities /28 native tangent evaluations maximum, zero optimizers.
- Four numerical threads; fifteen minutes including compilation and extraction.
- Preserve original data, assumptions, physical bounds, covariance and scores.
- Freeze point vectors, sources, parent receipts, environment and originals. Charge each unique point before calling native code; save atomic hashed full predictions and projected residual/Jacobian arrays under exclusive locks.
- Every center must reproduce its saved residual/Jacobian before its other points are evaluated. Keep failures and unused allowance; never retry or restart.
- Save scored horizontal and unscored vertical differences. Publication makes no integrations. No fit/repeat label, production gate or 53-stretch tally changes.
These are probes, not a new optimizer or a proof of a global optimum. A more favorable probe can justify a separately frozen recovery method; it cannot retroactively pass a failed repeat comparison.