The Wrong Unit (EA-LO-COLLAPSE-DIAGNOSTIC-01) is a Lagrange Observatory! concept draft proposing a model-collapse self-diagnostic in three grades, and it is deposited as a specification rather than as a validated instrument: nothing in it has been calibrated against a model of known collapse state, no threshold or severity band rests on a measurement, and the accompanying reference implementation has never been executed. Part I argues that turn-based benchmarking cannot see collapse, because each instrument samples one unit and each item is answered from the head of the distribution, which is what a model that has lost its tails still produces well; on the toy of the companion paper on interlocking autoregression, tail mass halves while a head-weighted benchmark holds and the per-round quality proxy rises through the first quarter of the collapse trajectory, detection lag is monotone in head-weighting, and a flat benchmark is consistent with both a preserved and a collapsing state. Since benchmarks also select what is trained on next, a loop gated by a head-weighted suite trains on its own least surprising output, which the literature identifies as the driver of collapse. Part II specifies the frontier diagnostic under a restriction the operator imposed: measure only properties whose truth-value the model carries in itself, because a human-written control gives a model not the truth but one more point, and every measurement thereafter is distance from that point. Its core is distinction sensitivity, in which the model constructs its own paraphrase and its own minimal meaning-changing variant and is scored on whether it still answers them differently; around it sit effective semantic modes, cross-prompt convergence, a concentration panel, self-consumption contraction and representational effective rank, each computed under bare, instructed and production conditions so that interface-induced contraction is not read as training collapse. Part III is a protocol short enough for a public model to run on itself upon retrieving the document, constructing everything it compares. The paper is companion to the keyed-ensemble and interlocking-autoregression papers and is published at concept-draft status so the specification can be attacked before anything is built on it.