The instrument: for each predicate, count prefixes that some shared continuation provably distinguishes — a Myhill–Nerode-style lower bound on the states of any deterministic machine consistent with the verdicts up to horizon L. Nothing here is canned: the bound is recomputed in your browser (the status line counts oracle evaluations). Regular predicates come out flat, counters linear, stacks doubling — press log scale and the doubling row becomes a straight line, the geometric signature of "the distinguishable prefixes are the reachable stack contents." This is the calibration table from §1; the same instrument, pointed at the model's internal directions, is what certifies the depth counter and grades the soft stack.