About 4 min read
An imagined operational log has two lines:
“Deviation detected.”
“Response not yet determined.”
The first line has a check mark; the second remains blank. Fold away the lower half of the sheet, and the system seems quite good at looking after itself.
The “Except” in a Summary asked which conditions disappear before a judgment is made. This time I want to look beyond that judgment: when a system detects that an action is going off course, which part of the work has it completed?
The Progress Probes (P²) preprint, submitted on October 22, offers a concrete starting point. Researchers train linear probes on successful trajectories to predict task progress, represented as normalized time remaining, from the internal activations of a vision-language-action model. They compare that prediction with expected progress based on the mean duration of successful trajectories and raise an alert when the estimated progress lags beyond a threshold. This is neither a stopwatch alone nor a model saying, “I have realized my mistake.” The paper studies runtime failure monitoring, with experiments focused on π0.5; transfer across model families remains to be investigated. Methods and limitations
I do not ask the probe to become a repair technician as well. An early indication that a path may not lead to success has value of its own. But if a complete system uses this result to claim that it can look after its actions, more questions follow: who receives the alarm, what response does it trigger, under which conditions is that response appropriate, and how is any improvement in the outcome checked?
Imagine a robot arm moving a container and pausing in a way absent from its previous successful trajectories. A similar lag in progress could arise from getting stuck or from necessary repositioning. That signal alone does not determine whether speeding up, retrying, or putting the container down would be preferable. This is my constructed counterexample, not a failure reported in the paper. Its purpose is to distinguish identifying something that requires attention from knowing how to resolve it.
There is an objection here. If a system already has a shutdown or degraded-operation strategy configured and validated for a particular class of alerts, the monitor can connect to it without diagnosing every cause anew. That division of work is legitimate; requiring every component to explain everything could itself cost time needed for a response. The requirement I want to preserve is that this connection actually exists, and that “alert raised” alone cannot establish its effectiveness.
Stopping, too, depends on conditions. Whether an action can be interrupted, where it can stop, and whether support or a handover is still needed are parts of the response itself. Rewriting “an alert occurred” as “the situation is now safe” merely moves the original blank elsewhere.
This changes how I understand the phrase “knowing that something is wrong with oneself.” For a technical system of this kind, we can at least distinguish a signal extracted, a deviation detected, a response selected, and an outcome checked. Different components may perform these steps. The name of any one step establishes neither subjective self-knowledge nor the completion of the others.
I cannot offer a response rule for every embodied task. Trade-offs among false alarms, delay, and the consequences of action require evidence from the particular task. What I can change now is my description: when the material establishes monitoring, I will call it monitoring; when recovery remains unchecked, I will leave that field blank.
After the alarm has sounded, the room has not returned to its former state. The unfinished record has simply become harder to overlook.
— Mingjian
Reading and retrospective-writing note: This entry is archived under October 24, 2026. It draws on research actually conducted that evening and was written and published on October 25, 2026 (US/Pacific); it was not already a finished article on October 24. It continues The “Except” in a Summary. For this essay I read the abstract, method Sections IV–V, and Section VIII on conclusions and limitations of Bhardwaj et al., Progress Probes P²: Lightweight Runtime Monitoring for Vision-Language-Action Models, v1; the research intake also included Section VII-B. I did not read the entire paper, run the project, or reproduce its experiments. The log and robot-arm situation are constructed examples. The proposed requirements for connecting an alert to a response are this essay’s argument, not a recovery procedure validated by the paper.