Local decisions.
Local authority.
The reference operates offline. Cloud availability does not sit in its decision path.
SHEAF-OS BY 2D AI LOCAL INTELLIGENCE / ACCOUNTABLE ACTION
Intelligence.
With the power
to act.
The operating layer for physical AI. Turn typed goals into checked action, connect every decision to its world state, and keep the history.
A working local robotics runtime. A much bigger thesis.
THE WORLD IS THE INTERFACE.FIG. 01
VERSIONED WORLD
CHECKED ACTION
X / Y / Z → STATE / AUTHORITY / HISTORY
01 / THE NUMBERS
Small footprint.
Serious operating discipline.
6.29ms
Median action boundary
41 steps. One changing-obstacle mission.
↗26versions
History recovered exactly
Recover the evidence. Reauthorize the action.
↗1,048
Stopping endpoints replayed
Independent readback of the nominal model.
↗4.37MB
Signed release archive
The runtime release. Without the lab.
↗Recorded in the macOS arm64 local AMR software reference. External runtime dependencies required.
Explore the results & methodology ↗02 / THE PLATFORM
INTELLIGENCE IS A PROPOSAL.EXECUTION IS A PERMISSION.
Autonomy needs
an operating
contract.
A good prediction is a beginning. A useful machine needs the entire loop.
Sheaf-OS brings perception, spatial state, bounded proposals and command authority into an inspectable runtime. The goal is more than the next move: know what justified it, whether it was acknowledged, and what survives a restart.
See how it connects ↘Local decisions.
Plan the stop.
Not just the move.
Rank admitted stopping endpoints using the actual motion lease and full braking trajectory in the nominal body model.
A world model
with a past.
Versioned observations, recorded command outcomes and exact promised history make the loop inspectable across runs.
03 / INSIDE THE LOOP
EXPLORE THE OPERATING LOGICMove. Refuse.
Recover. Repeat.
Useful autonomy includes the moments when a command should never leave the runtime.
SHEAF-OS / INTERACTIVE SCHEMATICLOCAL AMR REFERENCE
Admitted route Changed observationILLUSTRATIVE / NOT LIVE TELEMETRY
MISSION / COMPLETEDA destination.
A destination.
And a way back.
The changing-obstacle A/B/A software task completed three ordered goals through the real runtime and broker acknowledgment path.
41acknowledged steps
in the recorded mission
in the recorded mission
- Observe a changing scene
- Rank admitted stopping endpoints
- Authorize, acknowledge, record
OBSERVE → SELECT → AUTHORIZE → ACT → REMEMBERSCHEMATIC, NOT A LIVE RUNTIME
04 / THE ARCHITECTURE
SIX LAYERS. ONE ACCOUNTABLE LOOP.From a sensed world
to an authorized action.
Spatial state and execution authority belong in the same conversation. Sheaf-OS connects the evidence behind a proposal to the command it can actually admit.
01Perception
Timestamped observations
↓02Interconnect
Bounded, typed exchanges
↓03World model
Versioned spatial evidence
↓04Proposal compute
Finite candidate selection
↓05Trust & verification
Freshness, contracts, authority
↓06Actuation boundary
Broker and driver acknowledgment
↵THE SHEAF-OS THESIS
Context isn’t
background.
It’s part of the
contract.
A spatially indexed, versioned world model retains the state and provenance behind a decision. Typed contracts connect that state to bounded execution.
The architecture draws on local compatibility and restriction maps: a way to represent how local descriptions fit together. Its value is in an inspectable operating loop, not a claim that mathematics replaces sensing.
Read the technical overview ↗THE EDGE WE’RE BUILDING AROUND
A command path
with receipts.
Keep the observations.
Keep the version.
Keep the outcome.
Planning, memory and lifecycle control come together in a compact release. It is a foundation teams can inspect, exercise and build on.
Talk technical evaluation ↗9/9Functional gates closed
10Lifecycle cases accepted
78Focused & integration controls
Inside the release ↗05 / THE BIGGER THESIS
ONE OPERATING FABRIC. MANY FORMS.The world is bigger
than a chat window.
Physical intelligence deserves infrastructure of its own.
We’re building around a shared idea: local intelligence, explicit authority and persistent context should travel together. Mobile robotics is the first reference. Aerial systems, adaptive software and embodied devices define the expansion direction.
Mobile robotics
A local AMR runtime, exercised through useful tasks and recovery.
Aerial systems
Extend the shared runtime through body-specific contracts.
Adaptive software
Bring versioned context and accountable action to agent workflows.
Embodied devices
Explore local context and execution boundaries beyond the robot.
06 / BUILD THE NEXT LAYER
ENGINEERS. PARTNERS. INVESTORS.Bring your
hardest next move.
Evaluate the runtime. Explore an integration.
Back the infrastructure for physical AI.