What is settled, and where the frontier is.

Metric ablation

The field's implicit metric is the weakest one.

Detection practice operationalizes burstiness as the standard deviation of surprisal. We ablated five surprisal statistics on two sets: a hard 4-versus-4 hand-written dev set, and a 12-versus-12 corpus of public-domain literary prose against distilgpt2 output. The ranking is the result, and it is stable across both sets.

Standard deviation of surprisal, the metric the field runs on, is the weakest discriminator on the hard dev set (d = 0.59, medium). Local jumpiness (mean absolute consecutive surprisal difference) and mean surprisal separate human from machine far better. The coefficient of variation reverses sign on both sets, so flat text scores higher: we drop it. The corpus effect sizes are inflated by the easy human-versus-distilgpt2 contrast; the transferable finding is the ordering, not the magnitudes. This is what fixes the vector to B = [var(L), kurt(L), mean_surprisal, fluc_abs_diff, punct_entropy].

The in-generation controller

A boundary-steered sentence-length dial.

Token-level arms (a single activation vector, a best-of-N LoRA, GRPO) all hit the same wall: they bought variance by pushing the model off-distribution, degrading coherence, and never moved var(L) above its own noise. The fix was altitude. Steer only the sentence-ending punctuation logits toward a synthesized length plan, training-free, with the computable metric as a running discriminator. This controls var(L) at its native altitude, the boundary decision, instead of the token.

Under the paired long-form protocol the negative result forced, the dial moves. The first run, at lambda 8 over 300-token generations across 12 common-random-number pairs, gives a paired difference of +56.4 plus or minus 26.9 in realized var(L) between the low and high plans, clearing twice its standard error, while coherence is unchanged (mean surprisal 2.761 versus 2.770). Every prior arm had raised mean surprisal to buy variance; this one does not.

At spec-grade scale on gpt2-medium (lambda 8, 480-token generations, 30 sentences, 8 seeds, 4 dial levels, 48 matched pairs), expected var(L) rises monotonically with the dial, Spearman rho = 1.00, and coherence holds (drift about 0.13). The verdict statistic matters: the dial is a within-sample manipulation (low and high share the seed and prompt), so the right test is paired, and var(L) is heavy-tailed, so a single outlier generation swamps any mean-based effect size. That is exactly why Cohen's d wobbled between 0.59 and 0.87 across re-runs: it is the wrong statistic for this metric, an instance of the project's own noise theorem applied to its own verdict. Under the correct rank-based tests the dial is robustly controllable: 40 of 48 matched pairs respond in the commanded direction (sign test p = 3×10^-6, Wilcoxon signed-rank z = 4.72, p = 2×10^-6, mean effect 4.4 times its standard error). Content is preserved: cross-dial similarity (0.36) sits at or above the base model's own seed-to-seed floor (0.29), p = 0.002. The honest residual is magnitude, not existence: paired dz = 0.64 (bootstrap 95% CI 0.46 to 1.20) still straddles the 0.8 convention, so a clean magnitude headline wants a larger base; the controllability itself is established.

The GRPO arm

Before the boundary method broke through, GRPO trained a LoRA directly on the computable reward (no labels). The reward curve separates cleanly by target level: high-burstiness prompts hold near -0.57 to -1.08 across 110 steps while low-burstiness prompts sit around -4.7, so the reward distinguishes the two regimes. It did not, on its own, yield a reliable var(L) knob at distilgpt2 scale, the same noise wall the negative result documents. The boundary-steered dial is what finally moved the metric above noise while holding coherence.