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

Headline. The serial-decoding throughput basin observed across nine information- processing systems in Papers 1–6 is data-driven, not architectural or thermody- namic. The corrected basin width is ~1 bit per source byte (the historical τ ≈ 4. 16 was bits per BPE token — a tokenizer-packaging artifact, proven empirically in §3. 7). Across nine falsification experiments at 92M and 1. 2B parameters, no archi- tectural, thermodynamic, intrinsic-compression, loss-function, or scale-dependent hy- pothesis fires. The data-driven hypothesis survives, refined into the equation BPT ≈ sourceₑntropy − f (structuraldepth), and is verified at effect sizes up to Cohen’s d = 400. 81 (§3. 12, p = 2. 84×10−15). All blocking items from the internal adversarial review are resolved; cross-architecture generalization, non-Latin scripts, and alter- native quantizers remain open and are scoped under Paper 7. 1. Abstract. Papers 1–6 established that serial decoding throughput converges to τ ≈ 4. 16 ± 0. 19 bits per event across architectures, datasets, and scales (Whitmer 2026a– f). Paper 6 proposed the inherited-constraint hypothesis: AI models converge on ~4 BPT not because silicon demands it, but because biology authored the training data. We report four original experiments plus five follow-up experiments resolving all blocking items and extending the result to 1. 2 billion parameters. Under a unified evaluation harness, SYN-8 achieves 9. 06 BPT (8. 0 under random-offset training; BPSS*=8. 61±0. 12) —all exceeding twice the language basin. A 1. 2B-parameter model trained from scratch on SYN-8 extracts 8. 0 bits per source byte, identical to the 92M model, confirming the basin tracks data entropy across a 14× param- eter range. Three intermediate entropy corpora (SYN-5/6/7) show perfect linear tracking from H=5 through H=8, with no architectural attractor near 4 bits. A paired repeated-measures architecture comparison reveals a small transformer disadvantage (+0. 14 BPT, p=0. 029), not a basin-creating ceiling. INT4→INT3 is a universal quantization cliff across eight models. Silicon sits 1015. 7 –1018. 8 × above Landauer (corrected for context-length artifact). The key new finding: PCFG-8 data (8-bit entropy with hierarchical grammar) achieves 6. 59 BPT—between SYN-8 (~8–9) and natural language (~4). Three loss functions all converge near source entropy, confirming the basin is not a cross-entropy artifact. The refined equation: BPT ≈ sourceₑntropy − f (structuraldepth). A critical methodological finding: BPT is experimentally proven to be tokenizer- dependent. The same model on the same data produces BPT=8. 0 or BPT=3. 8 depending solely on tokenizer vocabulary size. Bits per source byte is the cor- rect tokenizer-independent metric and should replace BPT in cross-experiment comparisons. A comprehensive re-measurement of τ across 9 models (Pythia 70M–1. 4B, GPT-2 small–XL) and 2 corpora (WikiText-2, LAMBADA) gives τ ≈ 0. 85–1. 30 bits per source byte on WikiText-2 and 1. 15–1. 57 on LAMBADA — not the 4. 16 bits per token pre- viously reported. The basin is real; the number is ~1 bit/byte, scaling with model capacity.