A Multi-Listener Blind Panel on Long-Form Text-to-Speech Failure: Pre-registration

Pre-registration for a multi-listener blind listening panel on long-formtext-to-speech failure. Drafted 2026-10-24, before any participant data exist.At the time of writing, the study site is deployed but unpublished, the Prolificstudy is a draft, and the participant and response tables are empty. The panel asks whether, at a 60 s request, listeners are more likely to reportthat they could not repeat the message for AuK-Flash than for ZipVoice, and testsfour further hypotheses (H2-H5) covering a second TTS system, the reference-audioconfound, a manipulation check, and task validity. It registers five hypotheses,seven planned contrasts (C1-C7), eight exclusion rules, and the decision ruleunder which the main claim is withdrawn. This registration is posted before recruitment opens. Nothing in the registrationmay be edited once a result is known; corrections go in a dated addendum, not ina silent edit.