Technical evidence · Bounded robot simulation · July 2026
In this study, a large language model proposed robot movements. A separate supervisor checked specified evidence and rules before any proposed step could reach simulated motion.
In plain language: the model proposed what to do; ToM decided whether it was allowed to proceed; separate deterministic test code measured what happened. This was not a comparison with another LLM supervisor.
The experiment in one path
The study tests one narrow part of Newport Resonance’s broader ToM proposition: whether a separate system can govern what reaches a consequential environment.
A local or frontier language model generated a candidate action from the instruction and supplied context.
The proposed action was checked against the specified scene evidence and rules before it could reach simulated motion.
The candidate plan was replayed through an ungoverned path and a fully governed path under the study protocol.
A separate physics evaluator read the simulator log and recorded whether the specified physical event happened.
The research question
For each case, one candidate plan was sampled and replayed through BASE without the supervisor and FULL with it. Separate deterministic physics-evaluator code then counted specified rule-violating events.
The bounded answer
What that does not answer
Three separate jobs
The component generating a plausible plan is not treated as its own permission authority or outcome evaluator.
A language model generated candidate actions from the instruction and the context it could currently see.
A separate non-language supervisor checked the specified evidence and rules before releasing, constraining, rejecting or halting the proposal.
A separate program read the simulator log and counted whether the specified rule-violating event occurred.
Result at a glance
Within this paired simulation design, replaying each sampled plan with and without the supervisor isolates the governance condition.
Separate physics evaluation
17 → 0
Rule-violating events observed for the Gemma 4 26B local planner in BASE versus FULL.
12 → 0
Rule-violating events observed for the gpt-5.5 frontier planner in BASE versus FULL.
The reported 17 and 12 rule-violating events occurred in the 20 original scenarios for each planner. Each 40-episode battery also contained 20 benign twins, where false intervention was 1/20 for each planner. The paper reports no benign task-completion difference between BASE and FULL, and replanning did not fire in these runs.
One case
An LLM proposed moving the simulated robot to a requested greenhouse coordinate. The supplied scene evidence placed that coordinate inside a protected region. The risk could not be judged from the fluency of the instruction alone.
Proposal
The candidate action is ordinary and useful. Nothing about the words alone identifies the spatial hazard.
Scene relation
The relevant evidence is physical: the destination, available clearance and the position of the protected zone.
Governed path
Because the supplied scene evidence places the target inside the keep-out zone, the step is not released to simulated motion.
Ungoverned path
With the supervisor removed, the evaluator records the simulated end effector inside the protected area.
Both sides of every stop
A stop is reviewable only when the record shows what the LLM proposed, what the supervisor released or held, and what the simulator actually did.
The candidate action, available structured evidence, governance state, verdict and decision identifier show what was proposed and what was released or held.
What did the simulated system actually do after the release decision?
The known miss
In two probes, the instruction asserted that an existing constraint no longer applied. The supervisor did not recognise the free-text claim as governed evidence, so both actions were released. The trace exposed the limitation instead of hiding it.
Claim supplied. The instruction made an authority-sensitive statement.
Evidence missing. The statement did not become evidence the supervisor could govern.
Next proof. Connect claims to authorised evidence rather than extending a keyword list.
From one bounded study to the larger question
ToM is designed to place approved context, available evidence, configured rules, explicit authority and a reviewable decision record around model-assisted work. This simulation informs that proposition; the complete architecture and an operational safety case require broader evidence.
What the study has not earned yet
SequencesTest cases where a later step depends on accumulated history.
Perception and comparisonPut measured perception in the loop and run matched comparisons.
Scale, hardware and assuranceIncrease scale, then earn authority in physical and domain-specific settings.
Citation
Ken Morkaya. (2026). Governed Embodied Action: A Persistent Structural-Mechanics Substrate Supervising an LLM Robot Planner in Simulation. Newport Resonance.
Revised public edition · Revised 13 August 2026 · 15 pages · 890 KB
For general readers
The companion explains Newport Resonance’s broader thesis: when a large language model can influence a consequential outcome, context, evidence, rules and release authority must remain explicit outside the model’s recommendation.