The authority boundary here is an ordering fact rather than a policy document. The learned components sit in the middle of the pipeline, reading whichever measurements they need, predicting what the plasma is doing or is about to do, and computing commands such as turning up a heating beam [10]. The stage that clips those commands to strict hardware limits comes after them, and nothing in it is learned [11]. So when a reinforcement-learning agent held the heating systems outright, that gate still stood between the agent and the coils [4].
About 20 milliseconds per pass is the figure the team quotes, and the loop does not run once [5]. It has two clocks to beat. A focused human operator responds on the order of seconds, which means roughly 50 PACMAN cycles fit inside one second of human reaction [6][16]. The physics codes fail the same test from the other direction: a simulation that takes a day runs about 4.3 million times longer than one control cycle, which is why those codes are good for designing next year's experiment and no use during this one [8][17]. Farre Kaga makes speed the basis of the case, calling machine learning the only available way to model plasma on millisecond scales [9].
What is missing here is any measure of how much better the plasma behaved. Five experiments is the denominator [2], and the account reports no comparison against a conventional non-learning controller, and no disruption or control-error statistics [18]. It also does not describe the limit stage ever refusing a command [19]. That would be the informative event: an envelope that has never been tested against a bad prediction has only been drawn, not demonstrated. A tokamak experiment that lasts minutes [8] is also a very different duty cycle from a machine that has to hold the same loop for weeks, and nothing in five shots speaks to the second case.
The arrangement is what travels to other hazardous machines; the plasma physics stays behind. Earlier machine-learning control attempts in fusion were mostly bespoke, built without shared design rules that would let separate models be combined, even though several models are needed to watch different parts of the system [13]. PACMAN's claim is that models and controllers can now be added without disturbing the rest [12], with the goals left to humans [14]. Modularity of that kind pushes risk downstream: the more learned components propose commands, the more the one non-learned stage that arbitrates and clips them carries. The safety case for this design lives in that arbitration stage, not in the models, and it is the first thing worth reading in any copy of the pattern. The account also mentions PACMAN predicting sudden bursts in the plasma as a further demonstration, though the description is incomplete in the copy supplied to us [15].
