The problem
The cleanup model is small so it can run on your Mac in real time. That is the right trade for speed and privacy, but it means the model will now and then add a word you did not say or drop one you did. When the output is going straight into your editor, a single wrong word is a real problem, not a rounding error.
The earlier approach checked the whole dictation as one unit, which had a nasty failure mode: one bad micro-edit could throw away the cleanup of an entire long passage, and the longer you spoke, the more it drifted toward rejecting everything. The safety net replaced it.
Validate the edit, not the essay
The core move is to judge the change, not the whole output. The safety net finds exactly what the cleanup model changed between your raw transcript and the cleaned version, evaluates each change on its own, and rebuilds the result one sentence at a time. A doubtful edit reverts only its own sentence to your raw words. Every other sentence keeps its cleanup.
How a dictation gets checked
Each dictation runs through a short funnel, from cheap and certain to careful and rare.
- Line the two versions up. A difference algorithm aligns your raw transcript against the cleaned candidate and isolates each individual change. Because the two are usually almost identical, this stays fast.
- Classify each change. A fixed set of rules sorts each change by type. Obvious, safe edits (removing filler, merging a stutter, a hedge, a listed homophone, a self-correction) are accepted outright. Meaningful deletions and large rewrites are escalated for a closer look.
- Judge the doubtful ones. Only the genuinely ambiguous changes reach the one AI step, described below. Changes that simply drop content are reverted with no AI inference at all.
- Reassemble sentence by sentence. The final text is rebuilt one sentence at a time. Three mandatory checks run first, so a value flip hiding inside a small "safe" edit is still caught: numbers cannot change value or order, named entities and spelled-out acronyms must be kept, and a sentence cannot gain more than five new content words. Any sentence that fails is put back exactly as you said it, word for word.
The one AI-judged step
The only machine-learning part of the safety net is a small on-device model (about 127 MB) that reads the raw and cleaned text as a pair and answers one narrow question: does the cleaned version contradict what you said? It downloads quietly after your first launch and runs entirely on your Mac.
It weighs in only on the ambiguous, reworded cases, and on the cleanup path it looks for one thing: a contradiction. It is stricter about numbers than about ordinary wording, demanding near-certainty before it lets a change touch a number, because a number flipped from $1,500 to $15,000 is a different kind of mistake from a reworded phrase.
Safe by default
Your raw words are always the fallback. A reverted sentence comes back verbatim, so the worst case is never a blank and never an invented replacement. It is exactly what you said.
The design stays conservative even when things go wrong. If the small judge model has not finished downloading, or fails to run, the safety net does not wait or guess; it reverts every escalated edit to raw. And the whole check runs on a tight budget, roughly two seconds and at most eight of those AI checks for an entire dictation, so it never stalls you.
Privacy: counts, not words
The safety net keeps a small record so you can see what it did, but it never stores your dictated text in that record. It logs counts and category labels only, for example "14 edits applied, 1 of 7 sentences kept raw, judge 2 of 2." You can find that summary alongside each entry in your transcription history.
Known trade-offs
The safety net is deliberately narrow, and honest about it:
- It covers dictation cleanup. Command mode uses a separate, whole-output check rather than this per-sentence funnel.
- Text confidently detected as non-English mostly skips this funnel.
- The judge model is not always present, since it downloads after first launch. Until it is, the safety net simply stays conservative and reverts escalated edits.
- In evaluation, a small number of disfluency fragments (roughly 11 of 150 in one fixture set) are reverted when they could have been cleaned. That miss is always in the safe direction: it keeps your raw words.
The safety net is what validates the cleanup model and approves the swaps made by homophone correction. Together they are one system: the model proposes, and the safety net makes sure nothing reaches your text that you did not say.