I’m trying to make actuator-boundary safety claims in embodied AI reproducible rather than anecdotal.

Once a model (or ROS stack, or planner) can issue actuator commands, failures become motion. At that point, prompts and policy layers aren’t really a safety boundary — what matters is what reaches the motor.

I don’t think we have a shared, reproducible way to both:

• Enforce simple physical caps (velocity, acceleration, effort, position) independently of upstream logic
• Generate artifacts that a third party can rerun to verify that enforcement actually worked

I’ve been prototyping a small hardware interposer (“Sentinel”) that sits between controller and actuator. The hardware itself isn’t especially novel. What I care about more is the standard + proof workflow around it.

The rough structure looks like this:

1) A machine-readable safety contract (SSC)

Defines units, modes (e.g., development vs. field), stop behavior, and required evidence fields.

2) A conformance harness

Includes enforcement tests, malformed traffic handling, and fuzzing / anti-wedge robustness.

3) Evidence packs

Artifacts containing config + build ID, enforcement outcomes, wedge counts, and latency distributions (P50/P95/P99 within a declared envelope), plus a simple verifier.

The goal isn’t “this makes robots safe.” It’s narrower than that.

The goal is: if someone claims actuator-level enforcement works, there should be a reproducible way to test and audit that claim.

Right now, robotics experimentation is easy. But safety evidence at the actuation boundary is rarely standardized or shared in a way that different labs could compare.

One way to frame this is epistemic rather than engineering: if enforcement claims cannot be independently rerun using shared tests and shared artifacts, then “safety” remains largely narrative.

My main uncertainties are:

• Is this layer actually leverageful compared to perception or planning risks?
• Would enforcement create nuisance stops that incentivize bypass?
• Is “evidence pack” the right abstraction, or is there a better way to formalize reproducibility here?

I’ve put up a Manifund proposal to fund the first public milestone (SSC + harness + evidence tooling), but I’m more interested in whether this direction makes sense before pushing harder on it.

If you work on evaluation, governance, robotics, or safety standards, I’d appreciate thoughts — especially pointers to prior art I might be missing.

Happy to clarify technical details if helpful.