Validation

What is checked, against what, and where.

Claim	Check	Where
RNG and geometry unchanged, bit for bit	golden vectors frozen from the original JS prototype (Node 24); native and WASM must match	ctest golden_vectors
Math functions equal V8's	sampled arguments vs Math.* (100 per function)	ctest golden_vectors
Determinism under any query history, tiling, threads	world_checks (query order, cache size, 8 threads = 1 thread)	ctest world_checks
Microtubule paths	every point inside the cell, outside the nucleus, under the cytoplasm surface; z follows the local ceiling (middle third of it populated); few near-vertical steps; few turns tighter than the minimum radius; three scenarios	ctest world_checks.microtubules
Dye statistics	lattice angles and stagger, ring at 12.5 nm, tip at 24.5 nm, linker range, exponential ON/OFF means, geometric blink count, log-normal brightness	world_checks KineticsStats
Persistent sites	constant rate over 0-100 s and 10000 s, ON mean, window slicing	world_checks
Splat and FFT placement	against verbatim copies of the previous implementation; parallel = serial	ctest sr_render
Widefield engine	job format through a CPU reference host; GPU vs CPU (<= 3e-4 rms from fp16 spectra)	ctest widefield, tests/web/wf_gpu_check.mjs, wf_gpu_d3d11
Brightfield engine	weak phase grating (thin and 8 slices) vs first-order theory (~1e-5), empty field = 1, energy, serial = parallel; core optical volume: tiling, threads, nucleus volume = ellipsoids	ctest brightfield, world_checks, tests/web/bf_split_check.mjs
C++ chirp-Z PSF (ZernikePsf.cpp) vs webSMLM's JS / the Java class	0 / ~1e-12 relative L2	ctest zernike_psf, tools/psf_parity_check/
Chirp-Z PSF vs webSMLM's own chirp-Z	n <= 6 Zernikes, double-helix mask, depth shift: 0.0000% relative L2	tools/psf_parity_check/
Chirp-Z vs the old direct quadrature	0.22-0.29% relative L2	history in docs/dev/
PSF and noise models vs SMLM Challenge methodology [sage2019]	research comparison; the Challenge ground truth is a measured PSF table, so it gives no Zernike targets	docs/dev/vectorial-psf-step4-smlm-challenge-comparison.md
Adapter output unchanged by refactors	pixel hashes of fixed configurations	tools/adapter_pixel_hash.py

Known deviations

• Blink PSFs are cut (since 2026-10-07): at the default Renderer.Quality (Realistic) a blink's splat leaves out every camera pixel that would get less than 3e-6 of its photons (Fast: 1e-5; Exhaustive: none), without renormalizing. A movie summed over many frames lacks that far halo; compare with webSMLM, or study the PSF's wings, at Exhaustive. Optics, the halo cut.

• The cytoplasm mesh uses std::pow, which matches JS only on the libm Node was built with (differences below 1e-6 µm elsewhere).

• Brightfield: only cytoplasm, nucleus and microtubules make contrast (no nucleoli, vesicles, lipid droplets...: not simulated yet, kept out on purpose); scalar forward multislice, monochromatic, flat Koehler field; CellField only.
• Widefield: the Gaussian PSF (psf-model=0, not the default) ignores defocus; CellField only.
• Drift: the z drift of widefield and brightfield frames is interpolated between foci 10 nm apart; the background map does not move with the sample.
• Not yet verified by eye in every Micro-Manager configuration (headless pymmcore tests cover every property and the render path).