Smaller calibration boxes, with complete coverage
The completed broad-box comparison passes correctness but barely narrows the ranges. Its saved remainder floors are below the target on all 18 known-feasible axes. This study tests finer static-map bounds without recomputing native trajectories or changing physical assumptions.
Fixed partition and evidence
Reuse all 48 parents from the completed coupled-map prediction comparison. Split normalized mounting coordinates 0 and 1 at their fixed binary midpoint. The original eight yaw sectors remain unchanged. Skip fixed coordinates and coordinates with no representable interior midpoint. Keep closed shared boundaries. An independent exact construction must verify that all children cover their entire parent without dropping any of the 27 physical coordinates.
Use the same rule for all six generated candidate/processing groups. Freeze every child before evaluating its range; neither the known calibration nor future truth selects a split. There are at most four children per parent, 192 child identities total. Cells are pieces of one uncertainty calculation, not independent scientific observations.
Each child reconstructs the original source reference and keeps its radii, native expression, all parent constraints and all 78 coordinates. Add necessary complete-map relationships for the smaller physical box. Independently verify their exported rows before contraction. Run at most eight contraction passes, then six signed supports with the same independent endpoint audit as the parent. No baseline support audit is rerun.
After the child's range or empty proof is saved, determine whether it contains the sealed generated calibration. Check every containing child, including all closed cells sharing a boundary. An empty containing child is a failure. Never skip or replace failed, empty or unhelpful cells. The final six-group range is the union of all surviving children and original yaw sectors.
Finite allowance
At most 192 identities, partition audits, source setups, coupling audits, eight-pass contractions and post-range truth audits; 384 interval-map evaluations; 1,152 one-second LP proposals and 4,608 exact support checks. Each signed proposal reserves its two helper support evaluations plus one old-certificate reevaluation; each endpoint audit reserves six more. Retain the 1,024-row/78-column contraction cap. Four workers use one numerical thread each. Expected wall time is roughly 9–12 minutes; the cooperative ceiling is 960 seconds. No retries, allowance expansion, native predictions/bounds, original-aircraft reads or observed fits.
Every operation is durably reserved before execution. Charged interruptions remain failures without replacements. Freeze sources, environment, parent inputs, partitions, child identities and baseline targets before launch.
Gate
All registered child calculations and cover audits must pass. Every generated truth must remain in every containing child, and each group must have at least one containing child. Compute outward union widths and compare the same fixed twofold target on every axis with the original projected 7.2-second baseline. Report the immediate parent comparison as well. Partial improvements do not qualify observed predictions or full-flight uncertainty.