LPC + a memoryless coder plateaus far above the entropy rate whenever the prediction error is a fraction of a quantization step: the integer residual has already destroyed the fractional phase of the prediction, and whether it falls near a bin centre or a bin edge is worth 0.3-0.4 bits/sample. At sigma 5 with a 31-tap 300-2000 Hz bandpass, that is the whole distance between the 14.3x LPC+ANS plateau and the ~23x the rate allows. So add a tenth method that keeps it: the same Levinson predictor at full precision, each sample arithmetic-coded under a discretized Gaussian centred on the real-valued prediction. It reaches 21.7x on the same block, with the float32 coefficients charged and every reported size decoded and verified. Drop the hollow order-0 limit bars while here. They invited a comparison against R that does not hold -- they bound a per-sample coder on one prefiltered stream, not the process -- and the conditional coder sails past them, which read as a contradiction. The information they carried was really about coder overhead, so each coder now reports the model it codes against (ANS its histogram, the arithmetic coder its own predictive distribution) and a strip under the chart scores output against model: 1.4-1.9% for ANS, its symbol table plus arithmetic loss. That separates how well a coder does its job from how good its model was -- and both coders here lose about the same, so the 14x-to-22x gap is entirely the model.