Prepare the complete prediction comparison
The six-source comparison passes. Horizontal allowances improve in the three profiled-processing cases, while some fixed-zero baselines remain tighter. Keep every valid old bound. This is not the complete six-group width gate.
The source/domain preflight now passes in all 48 regions. The immutable subdomain transfer and cap checks below are complete. The joint adapter also passes its six-case known-answer qualification. The separately frozen complete comparison is running.
Preserve training information on the smaller domains
The saved-only ancestry audit checks all 48 original candidate/processing/yaw regions. Each retains every inherited 51-coordinate training constraint inside the 78-coordinate representation, with the same source, reference center and geographic frame. All 48 physical boxes narrow.
The existing ConditionedAcceleration.match accepts only its exact original
domain. Implement a separate immutable subdomain proof that verifies the child
box is contained in the parent's box and that every necessary parent training
constraint remains. Keep the parent's conservative acceleration bounds and
nonlinear caps; narrowing the domain does not require new LP proposals.
Bind the restriction to the source, reference, parent certificate and child
domain. Reject changed frames, missing rows, expanded boxes and altered columns.
Do not bypass the mismatch or substitute a physical-only acceleration prior.
Before an expensive comparison, check the initial-source cap certificate on every one of the 48 declared domains. The native qualification covers its fixed control matrix; it does not establish support for every new calibration box. Preserve any unsupported cases without cap expansion.
Qualify the joint prediction path
The separate joint-schema bridge passes local center-transfer, coefficient and rejection checks. Qualify its training-conditioned/subdomain path with a small known-answer training/prediction control before launching the full comparison. Retain all 51 native expression columns and all 78 enclosing coordinates.
Then freeze a complete comparison over every original candidate, processing case and yaw sector. Preserve all inherited training constraints, transfer center rounding outward, independently audit new training rows and future supports, and reveal generated truth only after each range is sealed. Show both fresh ranges and intersections with the qualified 192-box comparison. Every axis of every complete group must meet the original twofold-width target.
Cost the new finite start/bound/support allowance using the recorded 92–99 second globe bounds and 54–58 second disc bounds. Never reuse unused work from a completed study, repeat failures, omit unhelpful sectors or transfer a bound between different references. No new observed fits follow until this gate and the recording-scale/error-model qualifications pass.