RLSOK

THE DISPATCH THRESHOLD

Only approved robot policies reach real hardware.

Shadow-first release control for learned robot policies in ROS 2.

RLSOK binds the exact release content, controller configuration, approval identity, time-bounded Permit, runtime policy, and test evidence before the configured local ROS 2 controller path.

Requests that fail release checks do not reach the configured controller adapter.

An accessible spatial model of a content-bound release assembled from artifact, transform, contract, target, and authority layers. The default example contains a received content hash mismatch. The RLSOK Dispatch Threshold closes and the configured controller adapter remains unreached.

Blocked release at the RLSOK Dispatch ThresholdFive bound release layers align before a graphite dispatch boundary. A content hash mismatch closes the channel, leaving the configured controller adapter unreached.RELEASE ASSEMBLYRLSOK DISPATCH THRESHOLDCONFIGURED CONTROLLER ADAPTERNOT REACHED
RELEASE ASSEMBLY
RLSOK DISPATCH THRESHOLD
CONFIGURED CONTROLLER ADAPTERNOT REACHED
EXAMPLE DATANOT PRODUCTION TELEMETRYBLOCKED
release_id
rlsok-shadow-demo-001
approved_hash
sha256:8f4a2b9…d17e
received_hash
sha256:8f4a2c1…912a
device_id
ur5e-lab-03
controller_id
joint_trajectory_controller
decision
BLOCKED
reason
CONTENT_HASH_MISMATCH
configured_adapter
NOT_REACHED
hardwareSignalSent
false

INITIAL USER

For robotics teams moving learned policies from simulation into ROS 2 deployment workflows.

The initial fit is a team with learned or autonomous policies, multiple artifacts or configurations, and a need to reproduce release identity, approval, revocation, or expiry decisions before hardware connection. Not every ROS 2 project needs this boundary.

SIMULATION-TO-REAL WORKFLOW

REPRODUCIBLE RELEASE IDENTITY

DECISIONS INSPECTED BEFORE HARDWARE

CONTENT-BOUND RELEASE

A checkpoint is not a robot release.

A checkpoint contains trained model state. An executable release also depends on the artifacts, transformations, target identity, approval, Permit, and runtime policy required for local execution.

CHECKPOINT

model checkpointtrained model state

EXECUTABLE RELEASE

  • model checkpoint
  • content hash
  • normalizer
  • preprocessing
  • postprocessing
  • joint mapping
  • device identity
  • controller configuration
  • approval record
  • execution Permit
  • runtime policy

ONE CONTROLLED PATH

One release identity. One continuous decision path.

RLSOK provides a fail-closed reference execution path.
ExecSpec

Binds the content and configuration that define an executable release.

REFERENCE SHADOW MODE

Evaluate the proposal. Leave the controller path untouched.

Reference Shadow Mode evaluates policy proposals using the state inputs available to the local Gateway, without controller dispatch.

result
POLICY PROPOSAL EVALUATED
controller_dispatch
NOT ATTEMPTED
evidence
RECORDED
hardwareSignalSent
false

EVIDENCE

Record what actually happened.

RLSOK Evidence is hash-linked, verifiable, and tamper-evident. It records the release decision and the execution facts observed by the reference path.
EXAMPLE DATANOT PRODUCTION TELEMETRY

DECISION BLOCKEDThe mismatch resolved immediately to a blocked release decision.

release_id
rlsok-shadow-demo-001
approved_hash
sha256:8f4a2b9…d17e
received_hash
sha256:8f4a2c1…912a
Permit
permit-shadow-0042
device_id
ur5e-lab-03
controller_id
joint_trajectory_controller
decision
BLOCKED
reason
CONTENT_HASH_MISMATCH
hardwareSignalSent
false
previous_hash
sha256:10d9ab7…44c0
evidence_hash
sha256:d3c229e…a701

RESPONSIBILITY BOUNDARY

Release control is not functional safety.

RLSOK determines whether a specific release is eligible for the configured controller path. RLSOK does not determine whether the resulting physical motion is safe.

RLSOK CONTROLS

  • release identity
  • content binding
  • approval state
  • Permit validity
  • target device and controller checks
  • local dispatch checks
  • execution Evidence

REMAINS INDEPENDENT

  • E-stop
  • safety PLC
  • motion planner
  • certified controller
  • controller limits
  • mechanical safeguards
  • site operating procedures
  • independent hazard analysis

SHADOW EVALUATION

Evaluate one learned-policy release before connecting hardware.

Start with one learned policy in one simulated ROS 2 workflow and collect eligible and blocked Evidence without controller dispatch.
Request a Shadow EvaluationView on GitHubExplore the Architecture