Already Clean, Not Yet Meaningful: One "how" for the question Cammarata has opened

Cammarata's "Neural Networks Are Already Clean" argues that the canonical evidence for polysemanticity and superposition was misread, that what remains is recursive generalization, and that "polysemanticity is often more a claim about the observer than the neuron." The essay supplies a practice for looking but not an account of what the observer supplies when a neuron resolves, and it leaves "what you view as one thing" to the observer's judgment. This paper offers one such account, without claiming it is the only one. A neuron carries a cue, and a cue is a synecdoche whose meaning depends on habitus, context, and exformation; an explanation is adequate when it articulates the founding relation (Fundierung) between cue and substrate under those three. The neuron is therefore the right handle on mechanism and the wrong unit of meaning; the unit is the situated cue. Two formal sketches give the group claim a shape without establishing it: a Jensen-gap lens for what a one-number reading of a neuron discards, and a Gram-geometric description of what a small group of cues adds. On that basis the paper states, as a conjecture and not a result, that structure first appears when three cues are read together, that readings settle at roughly five to seven, and that larger groups become fragile. An appendix gives a test protocol and reports a small pilot on Cammarata's own InceptionV1 V-neuron figure, including an ablation that partly refuted the pilot's first reading and the re-reading that replaced it. What survives is modest: a group of upstream cues accounts for more of the target than any one cue, but two of the three named inputs do all of that work, no onset at three is visible, and the upper bound is untested. The reading that held under ablation is that one input carries a pointed pattern shared across categories and another is a weak category-specific supplement; the first reading, that the inputs were alternative routes, did not hold, and the paper reports that as an instance of its own method. The pilot also found that the figure's named vehicle input is a weaker instance of a vehicle-front route the figure does not name, which is offered as the paper's clearest evidence that reading with the quadruple finds what reading by top activations passes over. Nothing here claims the network understands; "clean" means legible to an anticipator, and the network does not care. The stance throughout is pragmatic constructivism, stated at the outset; readers holding a realism stronger than as-if realism will find the account unsatisfactory, and the paper says why.