Feature execution model

Where each capability actually runs, and what the current evidence covers. Read this before assuming a configuration field becomes a physical behaviour in the engine.

The badges#

Advanced pages in this documentation carry one or more of these. They describe execution ownership, not quality.

Badge	Meaning
`NATIVE`	Executed in the Unreal capture and runtime path.
`OFFLINE`	Applied after capture, by the Python pipeline, over the raw dump.
`PLANNER`	Represented or sampled by the planning and dry-run layer. Not by itself a physical engine behaviour.
`EXPERIMENTAL`	Implemented and reachable, but outside the currently verified public tier.
`EVIDENCE`	A measured or visually verified claim, tied to a recorded run.
`COMMERCIAL PREVIEW`	In the product today, with customer-facing proof still incomplete.

The boundary that explains everything else#

UNREAL ENGINE      owns the world and the raw truth
      |
      v
RAW DUMP           the durable boundary
      |
      v
PYTHON PIPELINE    derives, validates, reports and exports

Unreal decides what exists, where it stands, what the camera sees, and which pixel belongs to which object. Everything a trainer eventually consumes is computed from that afterwards.

The practical consequence: a different annotation format, split policy, report or validation rule is a re-derivation over an existing dump. A rerender is only required when the world, the camera or a capture modality changes.

The matrix#

Capability	Runs in	What that means today
Scene discovery	`NATIVE`	Unreal enumerates the level.
Scene helpers	`NATIVE`	Editor actors expressing spatial intent.
Spawn zones	`NATIVE`	Placement domains resolved in engine.
Surface placement	`NATIVE`
Grid placement	`NATIVE`
Spline placement	`NATIVE`
NavMesh placement	`NATIVE`	Projection-based placement onto navigation. Not uniform sampling over navmesh polygon area.
Asset pools	`NATIVE`	Compiled, then resolved in engine.
Cluster spawn	`NATIVE`
Relative / anchor spawn	`NATIVE`	Anchors come from actors already present in the captured scene.
Inter-class distance	`NATIVE`
Nested groups	`NATIVE`	Bounded recursive validation.
Spawn quantity distributions	`NATIVE`	For adopted population quantities.
Other recipe distributions	`PLANNER`	Sampled according to that field's execution path. A sampled value is not the same as a subsystem receiving a distribution object.
Actor representation	`NATIVE`
ISM / HISM representation	`NATIVE`	Identity-safe representation. Not a shared-batching performance optimisation; see the note below.
Camera sampling and zones	`NATIVE`
Fixed cameras and paths	`NATIVE`	Where adopted by the job.
Smart Viewpoint validation	`NATIVE` + `OFFLINE`	Native gate, Python threshold calibration.
Multi-camera	`NATIVE` `EVIDENCE`	Implemented capability. The latest dedicated visual QA run did not exercise a full multi-camera session.
Camera calibration / intrinsics	`NATIVE` `EVIDENCE`	Implemented. Latest visual QA exercises it only partially.
RGB	`NATIVE`
Instance-ID buffer	`NATIVE`	The identity pixels every label is derived from.
Depth	`NATIVE`	Optional, by capture and performance policy.
2D boxes	`OFFLINE`	Derived from recorded instance pixels.
Segmentation polygons	`OFFLINE`	Derived from recorded instance pixels.
Keypoints	`NATIVE` + `OFFLINE`	Unreal bones, Python projection. Only for configured pose-enabled classes.
Material randomization	`NATIVE`	Deterministic, preflighted, restored afterwards.
Animation randomization	`NATIVE`	Deterministic, preflighted, restored afterwards.
Dependency rules	`PLANNER`	Planning only. Native execution rejects them rather than silently ignoring them.
Ultra Dynamic Sky / Weather	`NATIVE` `EVIDENCE`	The current verified visual weather path.
Built-in NameFrame weather	`EXPERIMENTAL`	Not the current visual-proof path.
Lens, noise, compression	`OFFLINE`	Python image processing after capture, not a render-time camera model.
Occlusion target policy	`PLANNER`	A planning and QA contract. Actual occlusion comes from geometry and placement.
Verification	`OFFLINE`	Runs over the raw truth.
YOLO and COCO export	`OFFLINE`
Metadata v1 and analytics	`OFFLINE`
Coverage targets	`OFFLINE`	Measured report plus a draft correction plan.
REST API, SDK, MCP	`OFFLINE`	Local API v1 and governed adapters over it.
Release readiness	`EVIDENCE`	Gated on recorded evidence rather than assertion.
Commercial readiness	`COMMERCIAL PREVIEW`	A human and legal judgement. Not inferred from code.

The ISM and HISM note#

It is tempting to read ism and hism as performance modes. In the current implementation they are not.

All three representations preserve logical annotation identity, and measured benchmarking shows all three still carry per-logical-target component overhead. The instance-ID pass is a major scaling cost regardless of representation. Choose a representation for the identity and render contract you need, not as an unverified throughput promise.

Where the current evidence stops#

These are the boundaries worth knowing before you plan around a capability.

Unreal Engine 5.8 is the current verified tier. Other engine versions are not supported merely because the source might compile.
The dedicated visual QA path used Ultra Dynamic Sky and Weather. Built-in NameFrame weather is not the public visual-proof path.
Multi-camera and calibration exist in the product surface, but the latest dedicated visual QA did not run a full multi-camera calibration session.
Dependency rules remain planning-only and fail closed at native execution.
Lens and sensor degradation is an offline image-processing path, not a render-time sensor simulation.
Commercial Preview is not the same thing as customer-proven production readiness.

Stating these is deliberate. A capability list that never admits a boundary is a list you have to test yourself before you can trust any of it.

PreviousWhich path is mine?NextArchitecture