A complete orientation cover: generated feasibility plan
The joint-affine support component passes its independent software controls. The next experiment tests whether smaller native expansions improve the nonlinear remainder without dropping any original orientation possibility. This is a generated-data feasibility study, not an aircraft fit.
Fixed comparison
Reuse the input definitions of the completed six-case native contraction: three physical candidates, each under fixed-zero and profiled removal, 400 packets, the original 27-parameter bounds, known nonzero generated truth, six training epochs at 0.20, 0.35, 0.60, 0.90, 1.20 and 1.50 seconds, and a 1.80-second prediction. Use the predeclared 1 mm training allowance and the same 10 micrometre arithmetic allowance, which remains an assumption rather than a global numerical certificate. Do not import aircraft measurements.
Partition normalized initial yaw [-1,1] into eight closed dyadic sectors. Their union covers the entire original interval; keep every other parameter bound unchanged. Recenter within each sector, form necessary training inequalities with sector-wide remainders, and preserve their joint support. Use a validated box contraction only as an additional necessary constraint. Never discard a sector from an unverified LP infeasibility status.
Fix every center, training region and prediction bound before evaluating the future truth. Generated training observations may be used; future truth must not select sectors, widths, references or stopping rules. Retain every sector not excluded by a checked necessary contradiction, including sectors that contain no generated truth. Report the raw union, its nonlinear remainder widths, and the saved broad-box baseline without hiding worse outcomes.
Preparation and finite allowance
Before launch, implement and test exact coverage of the yaw interval, full-column retention, arbitrary sector centers, training/future separation, all native parameter/covariance identities and charge-before-evaluation behavior. Verify that prior generated inputs and baseline hashes match. Freeze the complete new implementation, generated inputs and numerical environment.
Proposed ceiling: 48 sector identities, at most four native evaluations and eight history bounds per identity: 192 native evaluations, 384 bounds and 153,600 reserved packet steps. Reserve all work before each operation. At most 288 one-thread LP proposals and 576 exact support certificates are permitted within those identities; each LP has a one-second cooperative limit. One worker uses four numerical threads, with serial LP proposals limited to one thread. Allow at most 600 cooperative seconds overall; the earlier measured bound cost suggests roughly six minutes. No retries, extensions, signals or reuse of another study's unused allowance.
Correctness requires the known generated parameters to survive their sector's necessary constraints and every future truth to remain in the final union. Every required case must complete. Report the width ratio for every axis and case; a tenfold reduction everywhere is a component-improvement target, not recording-scale predictive qualification. Correct but broad or incomplete results block dependent observed predictions and remain published.
The study does not establish actual measurement errors, a global floating-point certificate, long-record usefulness or absolute predictive adequacy. No observed Earth-model fit or protected ESA evaluation follows automatically.