Compiled multi-packet adapter qualification
Compare the compiled third-order interval adapter with the unchanged native tangent kernel. Preserve the first 64 packets, bit for bit, from each of the six qualified generated histories: three physical candidates and two processing assumptions. Preserve initial conditions, calibration scales and processing. Use normalized parameter value zero, with a query halfway through packet 64.
For each case, save one native position and all 27 first-derivative columns. Then integrate all 64 packets in each of the 27 unit parameter directions. Compare values and the corresponding first derivative using the original one-micrometre software-comparison tolerance and fixed coordinate scales. No column, including clock or removal, can be omitted. Higher derivatives retain their separate rational-reference and known-answer checks.
| Resource | Per case | Total maximum |
|---|---|---|
| Started identities | 1 | 6 |
| Native calls / conservative packet visits | 1 / 64 | 6 / 384 |
| Interval calls / packet steps | 27 / 1,728 | 162 / 10,368 |
| Value / first-derivative comparisons | 81 / 81 | 486 / 486 |
One worker uses one numerical thread, reserving two for the existing long-record study. The cooperative ceiling is 30 seconds per interval call and ten minutes overall. Freeze all sources, inputs, compiler and environment before launch. Charge before execution. Preserve failed starts; no retries, signals, allowance increases, aircraft-original reads or optimizer starts.
This is a short adapter comparison on stable packet/rotation branches, not uncertainty over the full calibration or clock domain. Passing does not qualify recording-length ranges, native execution-error bounds, complete branch coverage or an observational error law. Publication reads sealed results without reruns.