ADR-0008 — Research stays behind flags

Status: Accepted Date: 2026-09-22

Context

Virtual geometry, procedural microgeometry, representation virtualization, server meshing, and advanced GI are part of the long-term program. None of them has a measured win. Treating them as default architecture would lock the engine to untested designs.

Decision

Those five systems default off via FeatureFlags. Enabling a flag is an experiment. Promotion into the default path requires an ADR that cites recorded benchmarks and a fallback. RFCs define the success and failure criteria before the experiment is treated as a milestone exit.

"Better than Nanite" is not an assumption.

Alternatives Considered

• Build the experimental renderer first because it is the interesting part. Rejected. The founding promotion sequence starts with a correct ordinary renderer.
• Hide experiments behind comments instead of flags. Rejected. Dead code paths get enabled by accident.

Consequences

Derived-data keys include feature flags, so experimental products do not alias production products. Benchmarks live in docs/benchmarks/ and are never invented.

Supersedes

Nothing.

Superseded By

Nothing.