Full saved-point check of compensated coordinates
The small coordinate controls qualify compensated accumulation on six known-answer fixtures and retain the three-model compatibility checks. The next question is whether that arithmetic resolves derivative and stopping discrepancies in the same four exposed development examples.
Fixed work
Reconstruct the original inputs and error operator once per saved point. First evaluate the unchanged original implementation and require its baseline cost, parameter order and input hashes to reproduce exactly. Then switch only the diagnostic native provider to compensated coordinates with the qualified inward timing branch. Preserve the physical equations, bounds, error operator and original numerical environment.
For the compensated calculation:
- Repeat the residual and complete Jacobian three times.
- Compare every retained parameter column with feasible finite differences at the original normalized steps of 0.001, 0.0001 and 0.00001.
- Repeat the original undamped stopping probe once, using its original cost, stationarity and position limits. Record the signed actual and predicted cost changes and the parameter step.
The earlier matched diagnostic already isolated the timing effect. This phase checks the complete derivative matrix and stopping calculation together rather than reconstructing the same error operator for several narrower follow-ups.
Finite allowance
Four tasks, one per exposed point. Each permits one baseline evaluation plus at most 170 compensated residual/Jacobian evaluations: 171 per point, 684 overall. The complete planned probes need 650 evaluations if all finish.
The reconstruction cap is 23,634 single-packet operations per point. Including full evaluations, each task permits at most 23,805 native calls and 318,228,129 conservatively counted packet-query visits. Overall caps are 95,220 calls and 1,272,912,516 visits. Prepay separately; unused work remains charged. Do not reuse an earlier phase's allowance.
Use four workers with one numerical thread each, zero reserved threads only after other numerical work finishes, and a one-hour cooperative ceiling. Allow at most three compiler builds per worker: the original noise bridge, original wider tangent and compensated tangent. No retries or deadline signals.
Implementation and gate
Extend the existing saved-point study and reporting components with an explicit compensated/full-check configuration. Reuse their original reconstruction and source-bound probe substitution; do not copy another worker. Qualify the new configuration, complete-column probe, accounting and reporting before preparing its source/environment/input freeze.
Keep every failed original label. A passing new stopping diagnostic is not a new fitted optimum or a relabeling of the old result. No optimizer runs, forecasts against new observations, new noise families, aircraft reads or protected observations are included.
Publish all outcomes, including remaining disagreements. A failed comparison blocks dependent fitting. A passing comparison allows a separately frozen solver qualification; it does not restart calibration or establish a confidence threshold. The 53-stretch findings remain unchanged.