Long-record arithmetic and runtime comparison
The small prototype passes all nine source-bound checks. This separate gate tests six twenty-minute generated cases: all three physical candidates under both processing assumptions. It compares the existing wider arithmetic with the isolated mixed-precision derivative prototype at one fixed nonzero parameter vector. No optimizer or observed measurement is used.
Both methods receive identical 300 Hz increments, 1,200 query epochs, complete 26/27-column parameter domains and the same initial state. The trajectory contains turns and a climb. All positions are generated software inputs; this is not a training/withheld-prediction test.
Require bitwise-identical coordinates and pre-cast residual differences, including signed zeros, and identical timing-boundary counts. After dividing each derivative row by the fixed 0.05 m software coordinate weight, compare every column with relative and absolute tolerances both 2×10⁻¹⁰. Preserve any failure without relaxing these limits.
Measure process CPU and wall time for one native call per method in each case. Alternate the first method by processing case. Exclude construction, compilation and serialization from these kernel timings. A sampled speed ratio is not a guarantee about full fitting runtime. Preserve both complete value/derivative arrays, compiler/source bindings and binary hashes for saved-only verification.
The separate allowance is six identities, twelve tangent calls and 5,184,000,000 conservative packet/query work units. Each case constructs at most 360,000 intervals and generates 362,001 samples. Use one numerical thread and a cooperative two-hour ceiling, with no retries. Reserve capacity for the three other active workers. Charge work before execution.
Before launch, qualify the finite definition, strict value/column checks, closed-result guard and a real two-second pipeline bridge. Bind the complete source and environment set before its test report is produced. The earlier nine-check prototype evidence remains unchanged.
Use the shared study lifecycle and existing trajectory/problem/native helpers. Published verification reads saved matrices, metadata and journals; it does not repeat native calculations. Hash-identical binary artifacts can share archive storage. Source libraries are not copied to create another worker.
Passing this comparison still requires a separate full training/forecast qualification before any observed fitting change. The active noisy-data study, all existing outcomes and protected confirmation groups remain unchanged.