The genetic code as an information theorem: forced architecture, early freezing and the door of 20

The companion paper of this series showed what the standard genetic code is as a mathematical object; this one answers why it has this architecture. The answer is an information theorem, not a group-theory theorem. First, the code is exceptional against chance (0 of 2×10⁴ random codes dominate it simultaneously in chemical balance and robustness) but it is NOT the global or even a local optimum: it is an early freezing in the sense of Freeland, and living mitochondrial codes are still moving toward the Pareto frontier today — the two most frequent NCBI reassignments repair exactly the two impure blocks of the code. Second, the evolutionary chronology of amino acids follows the binary expansion tree of the code (Spearman ρ = 0.77, p = 8×10⁻⁵, robust across three independent chronologies); reassignment plasticity concentrates on the four wobble-block codons (37% of the 46 NCBI reassignments on a 6.2% base, binomial p = 1.0×10⁻⁹). Third, the forced-architecture theorem (T1–T5): given the chemical two-bit alphabet with antipodal involution and ~20 identity signals, any optimized molecular information system necessarily has words of length 3, 5-bit identity plus one free bit (wobble/punctuation), antipodal covariance (Hardy–Littlewood–Pólya rearrangement), reading hierarchy by information gain, and extreme error correction. The number 20 is attacked directly and closed with honesty: aminoacylation fidelity does not bound it (n_max ~ 600), the coverage-saturation program fails in both pre-registered metrics, and 20 is accepted as a declared input. The two routes to 20 (icosahedron faces, A4 roots) are proved to be a single A5-set, A5/Z3, whose color-Z3 fixed points yield one verified non-circular prediction: exactly 2 amino acids of degeneracy 1 (Met, Trp). Five no-gos are reported in full, including the theorem-level obstruction that A5 simplicity makes degeneracies 2, 3, 4 unreachable by orbits. Version 2 — correction notice. The reassignment p-value was overstated by ~6 orders of magnitude (binomial p = 1.0×10⁻⁹, not p < 10⁻¹⁵; a hypergeometric model is impossible here, as it caps successes at 4). The degeneracy comparison was stated backwards (ρ = −0.69 means high degeneracy does track antiquity). NCBI publishes 27 tables, not 21 (purity spans 0.88–1.00; Y spans 1.91–2.15; the standard code ranks 14 of 27). The X-optimum r_WC = −0.981 is now inside the stated Pareto range. A5 has five conjugacy classes and the permutation character is (20,0,0,0,2). The abstract now pairs ρ = 0.77 with its own p = 8×10⁻⁵.