2026-10-15. Native simulation cells and photograph pixels are separate budgets.
Four times the cells along each axis means 64 times the volume samples; eight times means 512 times. These studies initialize and evolve fresh fields at fixed physical extent. They do not enlarge a saved low-resolution volume.
| Process | Native grid | Physical duration | Evolution time | Evidence |
|---|---|---|---|---|
| Dose-owned thermal writing | 1152³ | 5; 3600 steps | 3414.43 s, Metal | 1.53 billion cells; three native fields independently scanned, finite and nonnegative. |
| Thermal writing | 768³ | 5; 1600 steps | 651.77 s, Metal | 453 million cells; all three native fields finite. |
| Thermal writing | 384³ | 5; 500 steps | 26.41 s, Metal | Finite native deposit, temperature and dose. |
| Acoustic deposition | 384³ | 6; 1500 steps | 207.75 s, Metal | Finite pressure, accumulated intensity and deposit. |
| Crossed acoustic fronts | 384³ | 2.4; 600 steps | 76.4 s, Metal | Off-axis pulses, driven source and mineral membrane; finite native fields. |
| Dielectric growth | 257³ | 800 growth events | 1001.88 s, Metal | 801 connected sites, 827 exact edges; final potential residual 4.371e−7. |
| Conserved phase | 384³ | 60; 240 steps | 825.7 s, CPU | Mean drift 5.48e−13; energy 64263→45650. |
Timings exclude compilation, allocation, export and photography. This shared
host slowed sustained GPU steps relative to short pilots; use complete-run
measurements when budgeting. Recipes live in experiments/refinement/; native
fields and run manifests in outputs/refinement-20261015/.
Writing at 96³ and 384³ uses the same path, width, extent, duration and dt=.01. Its volume-weighted deposit integral changes .0129232→.0129752 (+.402%). The .018 source width spans less than one cell per standard deviation at 96³, about 3.5 at 384³. Refinement resolves the writing process itself.
Phase at 96³ and 384³ uses identical physical modes and dt=.25. At t=60, matched-lattice composition differs by RMS .00794, maximum .0330; discrete energy differs .386%. These comparisons support this specimen's refinement; they do not establish convergence for every parameterization.
The shared phase engine now uses a real FFT half-spectrum. Splitting its inverse avoids a full-volume planning copy: at 192³, interleaved warmed steps measured 1.187→.643 s with about 14.5 KB allocated per step. The complete 96³ trajectory agrees with the old complex representation within 4.88e−17; 355 focused and coupled assertions pass. Old serialized phase checkpoints require their original source; new checkpoints preserve aliases and deterministic continuation.
The default World allocates eight fields and eight scratch volumes.
ThermalWriter needs two of those fields and owns three more arrays. The study
uses the existing parametric constructor to allocate exactly those five arrays:
9.06 GB at 768³, versus 34.43 GB for the default carrier plus writer.
The physics kernel is unchanged. A 120-step 96³ comparison is bitwise equal in nutrient, deposit, temperature, dose and clock. This specialized world supports the writer; processes or generic diagnostics requiring absent fields need a different allocation contract. Raw field exports are not executable checkpoints.
At 768³, explicit thermal diffusion requires a smaller timestep. The recipe uses dt=.003125, below .0033405, for 1600 steps to the same physical time 5. The complete run took 651.77 s of evolution (409.73 ms median step). Its three native field exports occupy 5.44 GB. The deposit integral is .0128622 versus .0129752 at 384³; this comparison changes both spatial and temporal resolution.
The old renderer multiplied a binary spherical crop into a scalar field before extracting its surface. That introduced artificial staircase rings. New photographs extract the actual field first, then cut the mesh with analytic geometry. A controlled 96³ comparison changes only that operation and removes the rings: this improvement is independent of simulation refinement.
Native 384³ phase meshes use declared .003–.01 surface-meshing adaptivity to remove redundant near-planar polygons. The scalar field is not downsampled or smoothed. Surface-only rendering loads the required native channel through a memory map, avoiding unrelated optical-array copies. The new target is 2400² photographs with 128 maximum adaptive Cycles samples.
Material presets change actual optical behavior: resin and quartz absorption; wax subsurface scattering; frosted transmission; thin-film interference; anisotropic bronze; copper with a chosen patina; porcelain and basalt. Patina is a shader choice, not simulated chemical growth. Transparent materials need sufficient optical thickness and suitable lighting to remain legible.
The old ecology dots were recorded agents, displayed as colored markers. New material views hide those markers and expose the actual paid film and deposits. The six diagnostic exports remain native 32³. Apparent connectivity changes strongly with the contour threshold; they do not show robust evolved anatomy.
Simply allocating 128³ would change the biology. A 32³/64³ resource probe kept total inventory and funding equal using cell-volume scaling, but an unchanged agent uptake request still changed .12025→.06050. Comparable refinement needs:
- Concentration-based sensing and saturation.
- Uptake and deposition over fixed physical supports.
- Physical interaction radii and offspring offsets.
The new cellular substrate implements concentration sensing, fixed physical exchange supports, continuous bodies and physical contact radii; comparable long-run spatial refinement still needs investigation. Older ecology recipes retain their original semantics. Mechanics stores continuous nodes and actual bonds; increasing a display voxel grid would not refine its degrees of freedom. The fresh 257³ dielectric specimen completed using residual-controlled GPU multigrid. Equal event counts at different spacings do not mean equal physical growth.
Seven reviewed 2400² refinements are published: wax phase section, resin writing, bronze calligraphy, copper arbor, crossed frosted wave fronts and patinated phase section, plus the native1152³ amber inclusion. The gallery now has 59 studies. The native 1152³ dose-writing run completed; its validated deposit integral is .01283649. Photography uses the original native field.