Diagnose the complete calibration map, not another matrix cut

The joint-rotation relaxation passes its software checks but barely changes complete widths. Preserve both completed freezes and their unchanged target.

The original 51 coordinates contain 27 physical calibration values and 24 arithmetic error terms. In a real calibration those terms are related functions of the same physical values; the current necessary region lets them vary more independently. Matrix inner products constrain only part of that relationship.

Question

For extreme points admitted by the latest region, keep their 27 physical values and recompute all 24 error terms from the existing source-bound calibration map. Does the resulting point still satisfy the training constraints, and how much does its affine future prediction change? This evaluates a generated calibration map, not an aircraft trajectory or a new native prediction.

Preserve every original coordinate. Report velocity/clock, lever, acceleration and force-matrix changes separately, but do not call these percentages of scientific evidence: their contributions can cancel.

Finite proposed diagnostic

Use the unique predeclared known-truth sector in each of six candidate/processing groups from the completed joint-product comparison. For each, propose the six signed affine extrema, with one-second LP allowances. Independently verify an exactly feasible witness using the qualified rational repair routine. Its known generated anchor may help repair the witness only in this explicitly diagnostic software experiment; it is not available as a prediction method.

Construct the 78-coordinate anchor by lifting the sealed 51-coordinate known point. Verify it against the current region. Record failed proposals, boundary collapse and unsupported rounding without replacements. Save the LP proposal, exact repaired witness and any physical-coordinate conversion before evaluating the calibration map. No source measurement or failed fit is repaired.

Validate six source-map controls at the known generated physical values against their sealed common-error values, using the existing 3e-10 component agreement criterion. Then make at most 36 additional calibration-map evaluations at the saved witnesses' physical coordinates. Reconstruct complete lifted coordinates and check every training inequality and box bound with exact arithmetic.

Ceiling: six identities / 36 LP proposals / 42 complete calibration-map evaluations / 250,000 scalar inequality-and-bound checks. Four workers use one thread each; 300 cooperative seconds, no retries or expanded allowance. There are zero new native predictions, integration bounds, original-aircraft reads or observed fits. Implementation must charge LP/map/check work before execution and pass focused accounting, source-binding, rounding and exact-feasibility controls before a new freeze is launched. This plan itself spends none of that work.

Interpretation

A physically mapped point violating the necessary training region cannot serve as a feasible witness for its prediction range. A passing point supplies only a feasible example for the surrogate calculation; it does not establish a global optimum, calibrated uncertainty or native trajectory coverage. Report both outcomes and the changes in every group, with no threshold tuning.

Use the result to choose a properly qualified nonlinear-feasibility or bound- propagation method. Do not continue a series of small linear-cut variations without evidence that they address the dominant representation gap. Other independent roadmap work, including ESA receiver-error qualification, remains available while observed prediction is gated.