Preregistration: Multi-listener blind panel on long-form (10-60 s) TTS failure, conducted on Credamo

This record is the preregistration of a multi-listener blind listening paneltesting whether long-form (10-60 s) text-to-speech failure is governed bygeneration duration rather than by the idiosyncrasies of any single system. It replaces the preregistration 10.5281/zenodo.23609599 (2026-10-24). Noparticipant was ever recruited under that registration and no data wereproduced; the panel described there was moved from Prolific to the Credamopanel. Design: 86 stimuli (84 test + 2 catch trials) in 14 cells of 6 stimuli each.The primary system is tested at 10, 20, 40 and 60 s on an RTX 4090D and at40 and 60 s on an RTX 5070 Ti laptop; three further systems are tested at 40and 60 s on the 4090D; a chunked-concatenation control and a non-distilledbase model are tested at 60 s. Stimuli are presented in 10 blocks of 30clips. 80 listeners (8 per block) each rate 28 test clips and 2 catch trials;every stimulus receives 24-32 ratings. Primary endpoint: the mean unintelligibility rate per cell - the average,over the cell's stimuli, of the proportion of listeners who could not repeatthe message. The proportion of stimuli judged unrepeatable by at least 80% oftheir listeners is reported as a secondary endpoint. H1 (the duration effect)and H2 (the effect holds across all four systems) are registered asdirectional hypothesis tests; H3 (chunked concatenation vs direct generation)and H4 (non-distilled vs distilled base) have only six stimuli per cell andare reported descriptively. Exclusion rules and the analysis plan are fixedin this document and will not be changed after data collection begins.