Benchmark matrix¶
This page is the concise source of current readiness. Numerical details remain in immutable JSON/HDF5 artifacts and device validation documents.
| Gate | Design 23 | Majumdar multilayer |
|---|---|---|
| analytical branch found and passive | pass: P0 → P1 at ~794–796 nm | pass |
| nonlinear residual recorded | pass | pass |
| complete geometry step executes | pass: 128 coordinates | pass: 75 coordinates |
| gradient audit | pass for dominant P3 smooth direction: implicit log-Q derivative agrees with exact P3 finite difference to 1.05%; native autograd.grad bridge tested |
historical phase checks; consolidated audit pending |
| reconstructed FDTD baseline | mesh 10 was unconverged; mesh 18 gives Q=11837.99 versus supplied Q=11812.39, with a remaining 4.06 nm wavelength offset | pass for Q ringdown; wavelength and full observables remain diagnostic |
| first direction confirmation | pass: calibrated AD candidate, Q +2.196% at mesh 18/full monitors | pass: phase 16 |
| second direction confirmation | pass: re-anchored AD candidate, Q +1.380%; +3.606% cumulative | pending |
| TE0 FDTD equivalence | full-monitor values available and valid; mesh convergence pending | pending; imported data is all-x flux only |
| FDTD convergence suite | Q ringdown passes at mesh 16→18; V and TE0 convergence pending | pending |
| supported status | no | no |
Next paid action¶
Design 23 needs no immediate paid action. Two P3-adjoint/multifidelity steps already passed same-mesh full-monitor checks. If optimization is extended, re-anchor at candidate 2, generate one bounded local proposal for free, inspect it, and buy only its identical-setting checkpoint. Do not resume the old raw-gradient trajectory or rerun duplicate geometries. A separate mesh sweep is still needed before mode volume or TE0 can receive nonzero objective weight.
For Majumdar, reuse the existing baseline/phase-16 artifacts, run five guarded steps, and buy only the next checkpoint. Require composite-objective and Q direction agreement, true TE0 monitor diagnostics, and loss closure.
Next free actions¶
- consolidate random-direction gradient sweeps into immutable fixtures;
- add adapter tests that translate legacy outputs into
EvaluationRecord; - run analytical truncation sweeps at the exact benchmark geometries;
- build FDTD convergence variants locally and estimate their costs;
- pin the Tidy3D serialization schema used by each recipe.