Engineering

Wrap, don't certify: runtime monitors and verification gates for physical AI

3 min readMati Melchior
Wrap, don't certify: runtime monitors and verification gates for physical AI

If you can't certify the model, what can you certify? The thing around it.

Yesterday's chapter showed why a text guard isn't enough. A VLA completes the task. The guard misses the implied harm. I found no VLA that claims a SIL under IEC 61508. So the 2026 answer is a wrapper: a check between the model and the motor.

Four recent sources show the pattern.

Check the action before it runs. Pre-VLA (arXiv 2605.22446, May 2026, preprint) scores each candidate action chunk before execution. It predicts a safety confidence and an advantage score. Low-quality chunks get filtered and resampled. On the LIBERO benchmark, average closed-loop success rose from 30.79% to 37.62%. Average verification time was 183.9 ms per action chunk.

Match the monitor to the model. SafeContract (arXiv 2605.28726, May 2026, preprint) ran three policy architectures through 450 episodes. Direction reversal predicted failure in all three. Velocity violations alone did not predict it anywhere. Yet the paper notes velocity checking is the most common safety mechanism in VLA deployment code. Its conclusion: "No single monitor works universally."

Keep safety out of the model's hands. In June 2026, Oxford's Fazl Barez called for "safety layers that don't depend on the AI being right". His examples: zones around people that a robot's arms cannot enter, and a physical emergency brake.

Watch from outside the robot. NVIDIA's Halos Outside-In Safety Blueprint, announced 22 June 2026, uses external cameras and AI agents to control robot behaviour in industrial settings. It's open source and in early access. NVIDIA says its inspection lab helps partners prepare for third-party certification. That's certification-ready. It isn't certified.

Here's how the four line up:

LayerExampleWhat it checks
Before executionPre-VLAeach candidate action chunk
During executionSafeContractmotor-command signals
Around the personBarezkeep-out zones, a physical brake
Outside the robotOutside-In Blueprintthe cell, through external cameras

One idea runs through all four. Treat the model's output as a draft. Check it before it becomes motion.

A note on related work. I published a method with the same shape, for a different job. The Verification Gate (a working paper, not peer-reviewed) is a work method for engineers who use generative AI tools. It puts one mandatory hold point between tool output and any commitment, such as a signature. It isn't a robot runtime monitor, and it reports no results yet.

Here's the honest limit. Pre-VLA and SafeContract are preprints. Pre-VLA reports benchmark success rates. It doesn't report harm prevented. SafeContract tested three architectures on two task sets. The Outside-In Blueprint is in early access. None of the four sources claims a certified monitor. A wrapper is only as safe as the evidence behind the wrapper.

Bottom line: wrap, don't certify = move the safety claim onto a part you can test. Then test that part like any other safety function, with a safe state it can always reach.

Safety engineers: in your safety case, which part carries the rating, the model or the check around it?

Part 5 of 12 · The Survey: what keeps physical AI safe in 2026.

Share

Physical AI Safety Dispatch

Monthly analysis. No spam. One exclusive insight per issue.

One issue per month. Unsubscribe in one click from any email. Privacy policy.

Regulation

What five weeks of counting found

104 days separate 8 October 2026 from 20 January 2027, when Regulation (EU) 2023/1230 will apply as a cut-off with no transition period. Thirty-nine notified bodies held a current notification on 13 September 2026, against one hundred and forty-four under the Directive it replaces. No harmonised standard has yet been cited under the Regulation, the application guide is expected in the fourth quarter of 2026, no readiness figure has been published, and none of the seven certification bodies checked publishes how long the work takes. The deadline is measured to the day. Almost nothing behind it is.

Read →

Regulation

You have 127 days. Nobody publishes how long this takes.

127 days separate 15 September 2026 from 20 January 2027, when Regulation (EU) 2023/1230 will apply. Seven certification bodies were checked for a published duration benchmark and not one publishes it, so what follows is what is documented instead: the first assessor notified on 8 September 2024, roughly twenty-eight months before the Regulation applies, no harmonised standard yet cited under the Regulation, and seven trade associations on record that machinery was the only sector left out of the Omnibus IV transition period.

Read →

We use cookies

This site uses essential cookies to function and, with your consent, analytics cookies (Google Analytics) to understand how the site is used. Learn more.