The Throughput Basin Origin: Four Orthogonal Experiments on Whether Serial Decoding Convergence Is Architectural, Thermodynamic, or Data-Driven

Six preceding papers established that serial decoding systems converge to approximately 3–6 bits per event and proposed that AI inherits this constraint through training data. This final paper distinguishes three competing hypotheses—architectural, thermodynamic, and data-driven—via four orthogonal experiments. Experiment 1: GPT-2 trained on controlled-entropy synthetic corpora achieves 8.92 bits-per-token on an 8-bit source, definitively falsifying the architectural ceiling hypothesis. Scaling to 1.2B parameters confirms scale invariance. Experiment 2: all eight models tested (Pythia and GPT-2 families) show a phase transition at INT4-to-INT3 quantization with greater than 190% degradation (a factor of roughly 3–14×), identifying the minimum viable weight precision. Experiment 3: transformers versus Mamba show no significant BPT difference (p = 0.688), consistent with architecture not being the limiting factor. Experiment 4: GPU efficiency is 1015–1018 above the Landauer limit, ruling out energy constraints. These results indicate that AI's approximately 4-bit throughput is best explained as data-driven: models learn training-distribution entropy, and natural language carries approximately 4 bits per token because it evolved inside biologically constrained brains. A critical finding is that BPT is tokenizer-dependent; bits per source byte is the correct metric.