Skip to content

Next Provider Selection RFC

This RFC records the provider selection decision for the next implementation batch. It is a planning artifact only: it does not change the provider catalog, generated provider docs, README, or public capability claims.

Decision date: 2026-04-30.

Selection Rules

Use the provider prioritization rubric before creating implementation issues. A provider is a good next candidate only when it has:

  • a stable callable surface that maps to one WorldForge capability;
  • a checkout-safe fixture strategy for success and failure cases;
  • a prepared-host smoke path with preserved evidence;
  • clear runtime ownership, licensing, and maintenance expectations;
  • no need for new base-package dependencies.

Provider names alone are not a reason to expand the catalog. The selected batch should improve real callable behavior or unlock a concrete host workflow.

The next batch should contain no more than these three provider additions or promotions.

Candidate Capability Owner Validation path Issue outline
JEPA-WMS score adapter score provider platform maintainer injected runtime tests, fixture-backed score outputs, runtime manifest, optional prepared-host smoke manifest Promote the direct-construction candidate only after upstream runtime loading, tensor shape validation, finite scores, JSON-native metadata, and failure typing are covered.
Genie interactive-world adapter generate media provider maintainer fail-closed scaffold tests first, then fixture-backed artifact generation and smoke manifest once a runtime/API contract is selected Replace the scaffold with one explicit upstream contract; document artifact type, prompt/state inputs, and why generated worlds are not planning proofs.
Spatial/3D scene adapter generate provider platform maintainer design issue, fixture schema for scene artifacts, artifact redaction tests, optional runtime manifest after contract selection Add a new issue before implementation that chooses one concrete scene/runtime API and defines the minimal JSON-native scene artifact boundary.

JEPA-WMS Issue Outline

Capability: score.

Owner: provider platform maintainer.

Validation path:

  • add a narrow adapter around the upstream JEPA-WMS scoring surface;
  • keep optional torch/checkpoint/runtime packages host-owned;
  • validate observation, goal, and candidate-action tensor shapes before calling the runtime;
  • validate finite score output, coherent best_index, and JSON-native metadata;
  • cover success, malformed score output, missing runtime, and missing checkpoint paths;
  • preserve an optional live-smoke manifest on prepared hosts.

Why now: it extends the JEPA-centered planning path without adding another media generator or generic provider name.

JEPA Public Adapter Addendum

Decision date: 2026-05-01.

Capability: score.

Selected upstream: facebookresearch/jepa-wms through torch.hub.load("facebookresearch/jepa-wms", model_name).

Decision: promote the public jepa catalog entry from a fail-closed reservation to an experimental score-only adapter that reuses the JEPA-WMS runtime contract. The adapter requires JEPA_MODEL_NAME, keeps PyTorch and checkpoints host-owned, and does not expose predict, embed, generation, transfer, or reasoning.

Migration: JEPA_MODEL_PATH was the old scaffold reservation variable. It is retained only as value-free diagnostic metadata; hosts must set JEPA_MODEL_NAME to load a real upstream model.

Why this does not promote every JEPA surface: upstream action-conditioned JEPA-WM planning maps cleanly to candidate-action scoring. Latent rollout and embedding APIs need separate contracts before WorldForge can advertise predict or embed.

Genie Issue Outline

Capability: generate.

Owner: media provider maintainer.

Validation path:

  • pick one concrete upstream runtime or hosted API before enabling capabilities;
  • keep the current scaffold fail-closed until that contract exists;
  • validate prompt/state inputs, generated artifact metadata, MIME/type hints, and artifact safety;
  • add fixtures for successful artifact creation and upstream failure payloads;
  • document the smoke command and preserved run manifest for prepared hosts.

Why now: Genie-style interactive-world generation is valuable, but only after the runtime boundary is concrete enough to test.

Spatial/3D Scene Issue Outline

Capability: generate.

Owner: provider platform maintainer.

Validation path:

  • write a design issue selecting one concrete scene or 3D-world API;
  • define a minimal scene artifact schema before adding provider code;
  • require fixtures for asset references, dimensions/units, safe metadata, and malformed payloads;
  • keep large assets, credentials, viewers, and GPU/runtime packages host-owned;
  • add a prepared-host smoke only after the artifact boundary is stable.

Why now: spatial artifacts are a distinct provider family and should not be squeezed into video or planning contracts.

Deferred Candidates

Candidate class Deferred reason
Additional remote video APIs Runway and Cosmos already cover the current remote media hardening path; add another only after it has a clearly different capability or user workflow.
General LLM reasoning providers The reason capability needs a stronger world-state question-answering contract before adding ordinary chat/completion adapters.
Simulator bridges The public scene/state boundary and host-owned simulator process model need a design record first.
New embodied policy stacks beyond LeRobot/GR00T Robotics validation cost is high; finish policy conformance, translator contracts, and prepared-host evidence before adding more policy runtimes.
Active inference or probabilistic belief adapters Useful direction, but WorldForge does not yet have typed belief/uncertainty result contracts.
More scaffold reservations Deferred by policy. A name without an executable runtime contract adds catalog noise.

Non-Goals

  • Do not change generated provider docs or provider README pages in this RFC.
  • Do not add provider names to auto-registration.
  • Do not add optional runtime packages to the base dependency set.
  • Do not present generated videos, scenes, or policies as evidence of physical fidelity without preserved run artifacts.

Follow-Up Rules

Each selected provider still needs its own implementation PR. That PR must state:

  • capability surface and implementation status;
  • runtime ownership and optional dependency boundary;
  • fixture and smoke evidence;
  • generated docs behavior;
  • redaction and artifact-retention expectations.

If a candidate cannot satisfy those requirements, keep it deferred and document the blocker rather than adding a scaffold.