Loudspeakers · 11
Directivity benchmarks and port-nonlinearity gates behind the CEA-2034 simulation pipeline. Method shown on hover; tolerance is the pass gate. back to QA
| Case | Tier | Reference | Gate |
|---|---|---|---|
speaker.independent-magnetic-enclosure-references.v1 | smoke | independent analytical magnetic and enclosure formulas | analytical agreement <= 1e-12 |
speaker.port-flow-promotion-contract.v1 | smoke | independent deterministic refinement-ladder contract | all observable changes <= 1e-2; at least 3 spatial and 2 temporal levels |
speaker.fem-bem-rom.v1 | physics-gate | full-order FEM/BEM solves, lumped Helmholtz resonance, and quasi-static compliance | sealed <=2e-2; quasi-static <=1e-3; coupled peak <=8e-2; curve <=1e-1; off-peak residual <=3e-2; damped <=1.5e-1 |
speaker.cuda-bounded-whole-workload.v1 | accelerator-gate | independent CPU f64 solve and field reconstruction plus CEA-2034 post-processing | CPU f64 true residual <= 1e-4; physical field error <= 1e-3; H2D/D2H evidence nonzero; all unsupported paths fail closed |
speaker.cuda-bounded-typed-driver-reference.v1 | accelerator-gate | independent CPU typed-driver solve and deterministic DriverModel reference artifact | residual <= 1e-6; power/velocity/impedance/derivative parity <= 1e-7; CUDA transfers and memory evidence nonzero; stale-grid and failed-promotion cases rejected |
speaker.cuda-bounded-typed-production.v1 | accelerator-gate | independent CPU f64 product and typed-driver parity/reference gates | residual <= 1e-4; physical <= 1e-3; derivative assembly <= 5e-3; coupled parity <= 2e-3; CUDA transfers and memory nonzero |
speaker.cuda-failure-contract.v1 | accelerator-gate | typed backend errors, CPU f64 acceptance policy, residency identity, preview bounds, and DriverModel promotion validation | failure_modes_exercised = 9; no fallback or skipped result is accepted |
speaker.native-x86-golden.v1 | accelerator-gate | deterministic schema-v3 artifact, portable DriverModel canonical bytes, and accepted aarch64 field/impedance/power/residual/gradient reference | all accepted aarch64 field/impedance/power/residual/gradient differences <= 1e-8; native target must be x86_64 and Rosetta is not accepted |
speaker.piston-directivity-benchmark.v1 | physics-gate | 96x384 overkill quadrature grid and closed-form 2J1/u directivity | self-error <= 1e-4; closed-form agreement <= 1e-2 |
speaker.port-nonlinear-2d.v1 | physics-gate | exact plane-Poiseuille profile, lumped coupled-flow estimate, and per-drive energy books | Poiseuille <= 4e-2; lumped limit <= 0.1; energy closure <= 4e-2; THD <= 1e-2; all steps complete |
speaker.port-optim-bounded.v1 | physics-gate | frozen Step-0 reference case; straight-tube baseline at equal voltage and equal airflow | Tier-0 feasible >= 6 incl straight; Tier-1 straight feasible; Tier-2 all healthy; SC1 pass; SC2 pareto-consistent; Q ratio >= 0.95; compression <= 1 dB; tuning <= 2 Hz |