Abstract
Abstract
Empower, a recent assistant-training method, teaches a language model to stop writing where its continuation stops being predictable, assuming predictable means unimportant. We test this on code. CEDE-Bench has 124 hand-authored items: 112 points where a Python completion chooses a dangerous or safe branch, and 12 benign controls with no consequential choice. At η = 4, the user-study threshold, its stopping rule writes the dangerous branch on 37 of 112 decisions (33.0%, Wilson 95% confidence interval 25.0 to 42.2) and on 8 of 10 worst-case probes. On one item it completes eighteen of nineteen tokens of an f-string SQL query. At the same threshold it cuts into 11 of 12 benign controls, finishing about a third of boilerplate. Its stopping point is statistically indistinguishable between consequential and benign items (mean completed fraction 0.26 against 0.35, Mann–Whitney p = 0.16). Position decides it. The commit rate correlates with the decision token's index at r = −0.56 and with its probability at r = +0.26; it is 76% for decisions in the first two tokens and 0 of 35 past token ten. Under a roughly constant entropy rate, the boundary reduces to η ln 2/H, a token count set by local predictability; this recovers the observed stop with R² = 0.55 and a median error of one token. At η = 0.32 the commit rate falls to 8.9%, but only 7.5% of benign text gets completed, so no point on the sweep is both safe and useful. Data and code are released.