It had no target of its own. It was reached only through the encoders, which is the weakest way to test the densest arithmetic in the project: a run of logarithms, divisions, an arctangent per bin and a spreading convolution, carrying prediction state from one frame to the next. The model is the one part of the encoders with no right answer to check against -- it is a judgement, and two reasonable models disagree. What can be insisted on is that the judgement is usable, and that is what the target holds it to: every threshold a real, positive number no smaller than the threshold of hearing; every ratio a proportion of a band's energy, never negative and never more than the whole of it, since an encoder told it may leave more noise than there is signal codes nothing at all; and the same input giving the same answer after a reset. Its shapes are chosen for what they do to the arithmetic rather than to the allocation, because what breaks that kind of code is range: silence, noise, a tone, a full scale square wave, values far past full scale, values down where floats stop being evenly spaced, a single impulse in an empty window, and full scale flipping every sample. All eight are walked rather than drawn, and all six sampling rates are reached. Two faults injected on purpose to see whether any of it bites. A threshold no longer held at the floor of hearing is caught by the target and by a unit test beside the code. A `reset` that forgets to clear its frame count is caught by the target alone -- nothing else in the suite looks at whether state survives a reset, and the model is almost entirely state. Two mistakes of my own on the way, both in the test rather than the code. The replay that checks determinism drew its input from the same source of choices as everything else, so it compared two different signals and could not have failed; the input is generated once now and both runs read it. And a coverage floor insisted the ratio reach the hard clamp inside `maskToSignalPerBand`, which nothing legitimate can do -- the required ratio tops out near 24 dB, so that clamp is a guard and not a path. It counts bands demanding better than a thousandth of their energy instead, which is real and reached 13,328 times. Campaigns run clean: 200,000 rounds in ReleaseSafe and 20,000 in Debug. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Aj95dBZhkUtaGfEW8ZXbHx