THE OPEN TRUST RUNTIME FOR PHYSICAL AI
The verifiable
trust runtime for
physical AI.
Connect intelligent models to physical and simulated embodiments. Author immutable assemblies, rehearse bounded sessions, and verify every action receipt.
Your lab. Zero install.
Run 1/2/3-link planar dynamics in a 3D browser viewport, inspect mass matrices and Jacobians, and export model state.
Open the 3D StudioSlow brain. Fast spine.
Explore local Mini Studio assembly authoring, 15 mm proxy snapping and numeric placement, plus separate bounded-session rehearsals.
Explore Central World & LabEvidence, by design.
Single-use capability tokens, revocation fencing, and Python/Rust receipt verifiers make every simulation or bench attempt inspectable.
Inspect the trust model01 / THE EXECUTION BOUNDARY
From intention
to evidence.
Intelligence can propose a move. The runtime defines what is allowed, the controller executes within those bounds, and the verifier checks the evidence.
- 01PLAN
Plan & author
Author a canonical
seedcore.assembly.v2graph or propose a bounded action chunk. - 02GATE
Authorize the scope
The PDP checks authority, limits, and freshness before issuing a single-use session token.
- 03ACT
Execute within bounds
Rehearse in host simulation with watchdog fencing. MCU execution remains behind hardware and firmware qualification gates.
- 04PROVE
Verify the evidence
Python and Rust verifiers check action-bound telemetry and quarantine missing or stale receipts.
// A bounded proposal for physical or simulated motion
const intent = {
profile: "fast-spine-v1",
assembly_digest: "sha256:2550d184…",
action: "joint.chunk",
bounds: {
max_velocity: 0.5,
horizon_ms: 500
},
authority: "single-use-session-token",
evidence: "action-bound"
};// Evidence verified by Python & Rust proof kernels
const receipt = {
intent_digest: "sha256:…",
capability_digest: "sha256:…",
telemetry: {
scope: "simulation_rehearsal",
action_bound: true
},
verification: "verified",
signature: "ed25519:…"
};Simulation tokens and proxy assemblies never grant physical hardware authority. Inspect the repository for active G0–G2 gates.
02 / SMALL SCALE. REAL ENGINEERING.
A foundation for your next build.
3-link & tabletop starter
Run 3-link planar dynamics in the browser and inspect the 5-part seedcore.assembly.v2 tabletop starter (0.2.0).
Microduck biped
Pinned 0.1.0 MJCF package, read-only microduck-sim-v1 HAL adapter, and 16 upstream-derived observation/ONNX goldens.
The fast spine
Bounded fast-spine-v1 simulation sessions, split-screen rehearsal, and the M0 / CW-D01 hardware candidate inventory.
Simulation & proxy authoring candidates are delivered locally · Hardware qualification (CW-D01, G1/G2) remains open.
03 / UNDER THE HOOD
Engineered across every layer.
C++ / Wasm ABI 2
78/78 deterministic tests.
1/2/3-link planar dynamics.
Central World v2
117 Python + 71 Node tests.
CW-A02/A03 local authoring.
Fast Spine v1
Bounded session rehearsal.
CW-D01 source inventory.
Rust / Python
Simulation token & receipt
proof verification.
YOUR NEXT EXPERIMENT STARTS HERE
Open a browser.
Explore the possibilities.
Start with the simulation. Inspect a model, run its dynamics, and follow the Central World and Microduck qualification gates.