Where Strata is headed.

Where we are now

Strata's voxel engine and bake pipeline are mature. A 99-module library is certified across three resolution tiers, from 4,096 up to 262,144 cells per module, with zero correctness drift. Backend selection is data-driven — the engine measures GPU versus CPU wall time per module and routes to whichever is faster, with no per-asset hand-tuning. Certification is per-asset and incremental, so the library can grow without re-proving what's already known-good.

The tooling around it is real, not aspirational: a cell-level inspector that x-rays individual voxel decisions, and an on-demand diagnostic exporter that produces a portable, human-readable archive of any bake run. The full numbers are on the technical page.

The path

1. Now

   Engine and bake pipeline certified at scale. Public site going up. Email list opening for launch notification.

2. Next
   • Editor lifecycle resilience. Bake pipelines participate in editor reload events so that script recompiles during an active bake cancel cleanly instead of stranding worker processes.
   • Worker protocol versioning. Workers refuse to run against an editor whose protocol version they do not recognize — a stale binary fails loudly, never produces subtly wrong output.
   • Automated unit test coverage on engine math. Symmetry-rotation invariants, voxel I/O round-trips, cert-matrix serialization, and backend-resolution truth tables become pre-commit gates — correctness regressions surface in milliseconds, not in a multi-hour certification sweep.
   • Faster developer iteration via assembly segmentation. Splitting the editor codebase into focused assemblies cuts the recompile-and-reload cycle for typical edits by an estimated 6-10x.
   • Coarse-to-fine certification scheduling. Cells provably clear at one resolution tier prime their children at the next tier — estimated 40-60% reduction in certification sweep cost for dense-interior modules. Pairs with the per-asset incremental model.
   • Cross-resolution symmetry priming. Symmetry analysis at one resolution warm-starts the next, eliminating redundant canonicalization work as voxel grids refine.
   • Diagnostic snapshot compact mode. A lightweight export tier emits only triage-critical fields, sized for fast handoff while the full archive remains opt-in.
   • Backend introspection observability. Engine pipeline emits structured per-phase telemetry through a stable contract; external dashboards can subscribe without parsing logs.
3. Later
   • Sparse voxel compression (DAG-based) for storage. As grid resolutions advance, on-disk voxel output transitions to compressed sparse representations that scale across orders of magnitude in cell count.
   • 128-cube resolution tier certification. Per-asset incremental, with coarse-to-fine priors from the 64-cube tier carrying as much certified context forward as possible. Requires worker-side enhancements landing in earlier roadmap items.
   • 256-cube tier and beyond. Requires sparse snapshot + sparse voxel persistence + streaming/chunked bake; the underlying data structures shift from full-grid to occupied-region representations.
   • Dirty-cell partial rebake. Only re-voxelize cells whose underlying geometry has actually changed, instead of full-module rebake on every edit. Estimated 30-80% rebake-time reduction depending on edit locality.
   • Parallel in-editor certification dispatch. Concurrent CPU bakes feeding a single GPU dispatch queue, with per-job lifecycle isolation. Targets ~3.5x speedup on the confirmation lane.
   • One-command module onboarding. A single action chains FBX import, module generation, face manifest, snapshot build, and incremental cert — adding the Nth module costs constant time regardless of library size.
   • GPU/CPU cooperative cancellation primitive. All long-running engine work polls a cancellation token at deterministic boundaries — the editor never has to hard-kill a running operation.
   • Strata as a standalone (project-extracted) engine. Long-horizon: refactor the certification kernel + voxel pipeline + symmetry subsystem into a reusable engine outside the editor's authoring tooling.

What we're not promising

No dates we can't stand behind, and no feature we haven't built. Strata ships when it's certified, the same way every module does. Follow the build and you'll hear it from us first.