Word lengths uniform in [2,10], letters uniform a–z.
Letters drawn by English frequency, first-order Markov vowel/consonant alternation (q→u, no triple repeats). Word lengths follow measured English distribution (mode 3, mean ≈ 4.7).
Bimodal word lengths (short mode + long mode) and boostable favorite letters.
Fixed 20-word line. Cross-entropy H(p,q) = −(1/N) Σ log₂ q(letter) per letter, where q = empirical letter distribution of each generator's last output (add-one smoothed to avoid log 0). Lower = generator's letters better match Shakespeare's.
H(shakespeare) = — entropy of the line's own letter distribution (plug-in estimate). Cross-entropy below can never beat this: H(p,q) ≥ H(p), gap = KL divergence.
| Generator | H(shakespeare, q) bits/letter |
|---|---|
| 1. Uniform | waiting for generation |
| 2. Markov | waiting for generation |
| 3. Dials | waiting for generation |