Physical source intervals at every packet prefix
The compiled transport and six-source runtime screen are complete. The new physical-source stream adds explicit calibration, rotation-center and matrix arithmetic allowances. Local checks include point domains, all 27 parameter directions, clock crossings, both packet phases and a nonexact mount matrix.
This qualification reuses the same six generated 1,600-packet sources, original
reference points and saved physical parameter bounds. It constructs physical
intervals for A = R f and B = R(P − I)f at every midpoint and endpoint.
It checks that the unchanged compiled source references lie inside the actual
reported allowances, with no extra comparison margin. It also checks the
norm-identity closure, including the literal installation matrix's metric defect.
This is a physical value calculation. It does not yet transfer the training acceleration constraints, qualify reference-row/Taylor-error conversion, install the body adapter or recompute complete position ranges. Those remain gates. The existing trained-adapter point failures remain preserved.
Finite allowance
- Six identities, six setups and no retries.
- Six complete source-value bounds; 19,200 midpoint/end bound steps.
- 9,600 compiled source-reference steps, 115,200 scalar comparisons and 19,200 norm closures, reserved before execution.
- Four workers, one numerical thread each; four total threads.
- 120 seconds per identity, checked cooperatively, and 600 seconds overall.
- Zero full native trajectory evaluations, LP proposals, optimizer starts, aircraft-original reads or protected ESA data.
Freeze sources, generated inputs, compiler and numerical environment. Record every reservation and failure. Compile numerical primitives from the frozen source directory. Keep interval/reference stream hashes and selected prefixes; do not store a recording-size derivative tensor. Every identity must complete and pass before the physical-source component is qualified.