The Quality Gate
The honest worry with any AI voice is a line that comes out fluent but wrong — a dropped clause, a clipped word, a stretch of dead air — and you only catch it three chapters later. Castwright runs two independent checks so that mistake gets caught before you do, not after.
Every rendered sentence is checked acoustically — dead air, near-silence, clipping, duration drift against what the line should take to say — before the chapter is assembled. A sentence that fails gets automatically re-recorded, up to a fixed retry budget, with no action from you.
A second, optional pass reads the words themselves. Off by default (turn it on in Advanced Configuration → QA gates), it transcribes each rendered line and checks it against the script — the "fluent but wrong" case the acoustic check can't see: a dropped clause, a swapped word, a line that says something other than what was written. Together the two checks cover both ways a take can go bad — the sound of it, and the sense of it.
A line that still doesn't clear after its retries ships anyway with the best take kept, and the chapter is marked Suspect so you know to take a listen rather than trust it blindly.
Expanding a Suspect chapter's row on the Generate screen shows exactly where the trouble is: a waveform strip with each flagged stretch rendered as an amber band, an "N issues to review" caption, and a tooltip on each band naming the reason — a line rendered suspiciously short against how long it should have taken to say, or a line whose words drifted from the script.
A chapter that clears cleanly shows none of this — no Suspect badge, no amber band, every character row reads Done straight through. That's the gate working quietly in the common case: nothing to review because nothing needed a re-record.
The acoustic gate catches a broken take. It doesn't catch a voice that rendered cleanly but drifted away from the character it's supposed to be. That's what the drift detector is for: once a chapter renders, it compares that chapter's synthesis against the character's established voice profile and flags any character whose rendered voice has wandered.
Flags are severity-tiered — Severe, Moderate, Mild — and open into a comparison view: the profile's voice attributes (gender, age, warmth, pace, authority, emotion) as they were "when rendered" against "now," a Listen control that A/B-plays the actual chapter audio against a fresh sample of the current profile so you can hear the drift rather than just read it, and a one-click Regenerate for that chapter. A Severe flag offers Auto-regen — no confirmation step, because at that severity the drift is confident enough not to need a second opinion. Anything you're not worried about, Dismiss (or Dismiss all) clears it.
The report opens scoped to the book you're in. If that book is part of a series, a This book / Series toggle sits at the top — flip it to pull in flags from every other title sharing the cast. However many chapters are flagged, the list loads a screenful at a time; scroll, and the rest follows.
A Suspect flag isn't stranded on the Generate screen. The same amber-marked waveform follows the audio wherever you play it — including the mini-player that pins to the bottom of every view. Hit preview on a chapter from Generate, or play it from the Listen tab's chapter list, and a bad take lights up amber in the scrubber before you've even pressed play — so a flagged line never has to be rediscovered by ear from scratch.
Everything above happens per line, per chapter — useful when you're chasing down one bad take, but not something you'd read start to finish for a 12-chapter book. The Quality gate card, on the Listen view (below the player, above the downloads) and on the Generate view (while a book is mid-render), rolls all of it into one card you can read in five seconds: a headline, four rows, and two export buttons.
The headline is blunt on purpose: "Every line held." if nothing needed a second take, "N lines needed a second take." if the acoustic gate caught and fixed something, or "Some lines need a second look." if the transcript check or the drift detector flagged something the acoustic gate can't see. Below it, four rows — one per signal this book actually generated:
- Acoustic — lines checked, and how many needed a second take (Check one, above).
- Transcript — lines verified against the script, and how many were flagged (the optional content check, off by default).
- Voice match — chapters scored for drift against each character's profile, and how many mismatches turned up (Check two, above).
- Cast continuity — how many config changes have happened since this book last rendered, tiered mild/moderate/severe.
Every row is honest about the difference between "clean" and "never checked." A book with no stochastic-voiced characters (an all-Kokoro cast, say) shows "No stochastic-voiced characters in this book — nothing for this check to do" on the Voice match row — that's not the same claim as "Not run for this book — flip on render-integrity checking to catch mismatches automatically," which means the check was available and simply wasn't switched on. And if voice-match scoring ran but failed for some chapters — an embedding failure, say — the row says so explicitly ("N of M eligible chapters scored, X couldn't be embedded") rather than quietly reporting the chapters that did succeed as if they were the whole book. The same rule holds for the Transcript row: "Not run for this book" only appears when transcript verification is genuinely off, never when it ran clean. Nothing on this card is allowed to look clean by omission.
Download as text saves a <book-title>-qa-report.txt file with the headline and all four rows, in plain sentences — the fastest way to hand a client or a narrator proof that a book cleared the gate. Download as JSON saves the full <book-title>-qa-report.json — every count and every mismatch behind the summary, for anyone who wants to script against it rather than read it.
Next: Listening & Revising.