Horizontal ridge-probe continuation qualification

Scope

The completed fixed-point study reproduced all twelve centers exactly. Its ridge probes passed all twelve unchanged local numerical limits; the original probes reproduced the four flat-model failures. This qualifies neither a fitting algorithm nor paired-start repeatability. Preserve the failed parent study.

Test a separately versioned candidate that changes only the local correction proposal to a column-scaled ridge penalty of 1e-6 on normalized displacement. The scored physical objective, all nuisance columns, bounds, clock and final stationarity/cost/physical-motion thresholds remain unchanged. The ridge term is not included in the model score. It does not identify weak calibration.

Eight predeclared identities

Continue all four flat-model parent endpoints and the two unsubtracted zero-field globe endpoints under the same generated observations and physical assumptions. Add two fresh opposed-start controls, unsubtracted globe-with-field and unsubtracted flat-disc, using the original full-calibration starts.

Each identity gets at most 200 tangent calls including fresh checks plus two diagnostic predictions: eight starts /1,616 native calls maximum. Four numerical threads, thirty minutes, no retries or expansion. No original recording reads or observed-data fits. Interrupted identities remain charged.

Acceptance

Require all eight individual controls to pass the unchanged objective/error checks. Reuse only the six explicitly named passing globe companions from the parent, yielding exactly twelve primary comparisons and all six paired-start checks. All must pass, including the previously disagreeing zero-field globe pair. No outcome-based companion substitutions or relaxed repeat limits.

This study tests an explicitly staged continuation method, not a claim that the earlier one-stage allowance succeeded. Keep all old fit and qualification labels. Any subsequent observed study needs its own finite freeze and matched repeats.

Freeze source, input, environment and exact identity hashes before starts; record each native call before execution, preserve atomic outputs and use exclusive launch/worker locks. Package-only performs no numerical work. Never restart a charged or completed batch or borrow another study's unused allowance.