This public shelf covers engine-neutral evaluation and bounded physical-AI development. The host runtime remains authoritative; the model, planner, or evaluator can observe, propose, and be measured without silently mutating the world.
Simulation is a development and evaluation surface, not proof of hardware capability. A replay, mock run, or MuJoCo result can demonstrate a bounded contract while leaving real hardware, trained-model performance, and deployment readiness as separate questions.
One holds the evaluation boundary; the other explores a bounded simulated physical-AI development loop.
The harness keeps observations, model decisions, policy, replay, and evaluation separate from the authoritative world runtime. It is designed for long-horizon evaluation under partial information.
It does not replace a game engine, own the world state, or turn model output into authority.
A bounded development surface for objective-to-capability-graph planning, simulated bimanual table operations, vision, routing, verification, and evidence receipts.
The public evidence covers mock mode and simulation. Hardware performance, trained-VLA capability, and universal household behavior remain outside the claim.
The public development story is strongest when every step has an owner and an evidence tier.
The system receives an explicit observation or fixture instead of an unexamined world snapshot.
A model or planner produces a candidate action, capability graph, or explanation—not authority.
Policy, environment rules, and the selected host runtime validate whether an action can be applied.
Receipts, replay, and outcome analysis test the chain without upgrading a simulation into a deployment claim.
This page does not claim a proprietary public game engine, a trained household robot, a hardware-ready controller, or a production deployment. It describes the public development and evaluation surfaces that are actually exposed.
