Two design mindsets, one LLM policy, and the reason the new rules cancel each other

The comparison arrives lopsided, and the lopsidedness is the part that transfers to your org chart. The control family comes with borrowed vocabulary from four separate fields ^[9], with a theory of interfaces where deciding how much mess is tolerated inside a process is exactly what lets you publish a clean boundary outside it ^[12], and with software as its ideal case: hard-won determinism, same execution every time, no wear and tear ^[13]. The other family, in the text at hand, gets one clause about finding interactions and mechanisms that foster desirable emergent behaviour ^[4]. Six illustrative anchors against one description ^[17].

That ratio is roughly the ratio available to whoever is drafting the model-use policy this quarter. Decomposition supplies ready phrasing for tolerances, redundancy, and taking the broken thing apart to fix it ^[8]. Emergence-oriented design has to be argued from scratch in the room. So the artifacts come out control-shaped, and the corollary rides along unexamined: if every part plays its role well, the machine ought to work well ^[7].

The premise under that corollary is what gives way. Control requires taming a messy world and bounding variability ^[8], and software qualified as the ideal case because the determinism was already paid for ^[13]. Hebert's observation that humans and LLMs are not interchangeable ^[2][16] lands here rather than on quality: you can keep per-component acceptance criteria intact while the thing producing the components stops honouring the premise those criteria rest on. The ensemble then misbehaves with every part having passed its own check, and the procedure that follows is the decomposition one, dive in, find what broke, fix it, be better for it ^[8]. Correct for a machine. A category error where the behaviour sits in the interactions ^[4].

The human factors material is the sharpest instance. Ideal observers, meaning a theoretical person monitoring instruments at optimal frequency, exist so that complacency can be defined, and working memory limits exist so that systems keep the people inside them acting within desirable parameters ^[10]. Any policy whose whole safety argument is that an engineer reviews everything the model emits is leaning on that construct and inherits its ceiling. It is a control-side answer, and it does not get stronger when a second policy in the same handbook asks the same engineer to ship more.

None of this requires a view on whether the models are any good. The instruction is upstream of tooling: systems are systems, and the known patterns are available as guidance and warnings rather than as novelty ^[14]. Teasing apart assumptions has always been the job of this comparison ^[15], and the assumption most shops are holding is that determinism came with the language rather than with whatever is writing the code ^[13].