July 2026 prompt evidence

Each detector returns a human-likelihood on a common 0 to 1 scale, where 1 means it judged the text human-written and 0 means it flagged it as AI. On these pages a verdict counts as passed when that score is at least 0.50, the midpoint of the detector's own scale. That threshold exists only to draw the chips: the bypass rate on the leaderboard is the mean of each test's median score across the 5 detectors, a continuous number, so a tool's pass count and its bypass rate will not be the same figure.

A few runs come back with fewer than 5 verdicts because a detector did not return a usable score for them. A run still counts as passing at least 4 detectors only if that many detectors actually scored it and passed it. A partially scored run is counted as caught if any detector that did score it flagged the output, and is otherwise listed separately as partly scored.

Meaning is the input↔output embedding cosine and readability is a language-model writing-quality rating, both published per test in tests.json. Words is the output's length as a multiple of the input's; the scoring code penalizes ratios above 1.40 or below 0.60. Full definitions live in the methodology.