[purinda]

Summary

Speech-to-text writes a full stop where the speaker paused, and capitalises letters spelled out and words it mishears. Deep's check (SelfRepair, in Swift and Rust) took both as the speaker's own, so it turned down right answers to real transcripts, and blocked many measured training pairs. Now it reads them as guesses:

• Two-item lists. A full stop sets two things off as a colon does, and so does a comma when the lead counts them ("two things, …"). The Deep training validator no longer asks for a colon.
• Corrections across a full stop. A cue that opens a sentence may correct the phrase the sentence before ends with ("…in the garage. Actually, the lobby." → "…in the lobby."). It may not take back a whole sentence ("I finished the report. Sorry, I was late." stays), a new clause, or "two chairs" with a bare "three". Corrected words said again may end their sentence ("not Sunday. Thursday.").
• Capitals. Letters spelled out may be joined ("p r" → "PR"). A capitalised word may be respelled where the capital says nothing: written without one, or where a list item starts ("First, Madge the PR" → "1. Merge the PR"). A capital still guards a name everywhere else. Two words are never merged across a sentence end ("plan A. I think" is not "plan AI think").

docs/cleanup.md, design-notes.md and limitations.md describe the new behaviour and its limits.

Guard fixtures

• Fixtures/cleanup/guard.jsonl has 64 new cases. Most lines change because each said word now records startsSentence.
• Three earlier verdicts change, each to accepting the right answer:
  • "Paint the door red. Actually, blue." → "Paint the door blue."
  • two "Meet on Monday. Sorry, Tuesday. Bring two chairs. No, three. …" cases that had hit the limit of two corrections that go back. A new case keeps that limit: three such corrections stay rejected.

Test plan

• Swift CleanupTests pass (xcodebuild test).
• cargo test -p lt-cleanup: 195 pass, including the guard fixture replay.
• cargo clippy -p lt-cleanup --all-targets -- -D warnings and cargo fmt --all -- --check are clean.

Not in this PR

• The Deep adapter retraining on measured speech-to-text data, which is still being evaluated.
• Gaps in the check found while measuring new adapters. Main has them too, and they're tracked separately.

🤖 Generated with Claude Code

[purinda]

Superseded by #58, which merged this work into main with the other changes behind candidate E. It shipped in 1.2.0.