1. Two musicians can agree that they have settled while disagreeing about how they got there
They found: Paired annotations agreed strongly once a stable musical space had formed—median stability overlap was 81.3%—but much less on proposals (23.6%) and almost never on negotiation (median 0%; 27 of 37 clips had no overlap). The d…
2. Making text easier is unsafe if the answer quietly disappears
They found: Against same-backbone controls, DFT-GEN won 93.3% of English and 64.2% of Chinese rendered-accessibility comparisons. In the pilot it kept correctness near 94–96%, lowered reported effort to 1.3/7, and improved completion tim…
3. A coding loop is accountable only if every repair remains attached to an owner-approved claim
They found: The prototype can bind a reported problem to a target interface state, reproduce it, route it into a revision, retest only the affected claims, and return a candidate to the owner for release. The contribution is the traceabl…
4. Chat did not beat search—or harm immediate learning—on familiar debates
They found: There was no significant difference between chat and search in argument expansion or critical reflection. Chat users spent more time with the tool, while exploratory patterns suggested prior attitude strength and intellectual…
5. A neural instrument can expose the model’s awkwardness as playable material
They found: Stronger latent regularization made continuous movement more predictable but traded away some corpus structure. Smoothing turned twitchy micro-temporal controls into slower gestural contours. Imperfect reconstruction, spectra…