A Workaround With a Measured Reason
Two lessons ago the take pipeline started every take warm. The first take got a plain, neutral cue in front of it; every later take borrowed the last cues of the take before it. The warm-up rode in the text sent for synthesis and in the cache key, and a cut located by transcribing the audio removed it before anyone heard it. None of that was decoration. On the older voice model a take that opened cold could read its whole first stretch in a stiff, robotic tone, and a warm-up that sounded like a broadcast greeting made the model read the entire take like an intro; each of those had been measured. The machinery built around them — the opening cue, the borrowed tail, the boundary finder, the trim and re-cut — was a careful answer to a real defect.
It also never fully went away. Slivers of the warm-up survived the cut on several takes, a transcription check once reported an opening clean while a leftover word was still audible, and in the reference episode the only catch at the final watch was exactly such a word.
The Premise Changed Under It
At the end of September the default voice model moved: new takes now inherit the voice engine's per-profile choice, and Pippa's voice runs on the newer model. The first episode made on it was rendered twice. Version A primed every take as before. Version B sent every take as nothing but its own authored text. The owner listened to both and ruled on 2026-09-29: B had no residual warm-up sounds, and the tone did not change from one take to the next. He had said earlier that the warm-up would go once that condition was met, and it had been.
The retirement was complete rather than partial. The opening-cue generator, the borrowed-tail resolver, the boundary finder, and the trim-and-re-cut code were removed. One resolver now builds the text of every take for synthesis, cache identity, timeline lookup, declared edits, the pace report, and an exact reroll, so there is no second place where a warm-up could quietly come back. The old priming fields in a script are refused before any paid synthesis or cache change, so a stale script cannot re-enable the retired path by accident. What stayed is everything that still answers a live defect: breath restoration, alignment, the coverage gate, the runaway-tail repair, model receipts, and the independent check of every take's opening in the finished movie.
The productions that were primed are not reinterpreted. Their sealed toolchains replay them exactly as they were approved, and the last implementation of the warm-up remains in version history. The pace thresholds measured on the older model stay in the report as prompts for where to listen, never as verdicts on the newer one: the owner approved B's delivery, and a number does not overrule that.
The Same Fortnight, the Same Shape: Numbers
A second workaround fell the same way. The script used to spell numbers out in Korean words for the voice, so the model would not have to guess how to read a digit. On 2026-10-10 a travel script turned the owner's dictated "7 minutes" into a Korean phrase that means seven people. A number spelled out fixes one reading, and no unit table can undo it. The newer model reads a digit by its meaning, so the rule reversed: numbers stay in digits, as they were said. Counts that Korean naturally writes as words stay words. The gate that enforces it does not keep its own patterns; it asks the voice engine which numbers in the script have been spelled out, and an unreachable voice engine fails the check rather than passing it.
A smaller fix from the same weeks carries the same lesson about evidence. Subtitle timing is anchored by matching each cue against the provider's character clock. When the voice engine began rewriting some text before synthesis, the script's own words stopped matching what had actually been spoken, and cues lost their anchors. The match now uses the line exactly as the voice was sent it.