v0.1.14 — Typing over the in-browser model, measured at a human pace

Typing over the in-browser model stays smooth now, measured at a human pace, and rewrites keep quotations and words in another script.

Typing over the in-browser model, measured properly

0.1.13 stopped a live check when typing resumed. Measured again at a human pace (one key every 80 ms, a 2 s pause after each sentence) instead of in bursts, that did not hold. Each check spent 3.7 s reading ProofKey's check prompt before writing anything, a stretch the model cannot be stopped in, and the check sent again at the next pause paid it again: in Brave on an RX 6600, 33 to 121 of 462 keystrokes took over 100 ms to paint.

Live checks now continue one conversation, so the model reuses what it already read: 0.29 s to its first word instead of 3.7 s, and 0 slow keystrokes in every run after warm-up, in Brave and in Edge. The stop-on-typing is gone; it has nothing left to stop. On the 14 live-check fixtures the quality held (258/280 against 260/280 sent fresh, no false alarms either way).

The model is now also loaded, and the conversation started, when a page with live checking opens, rather than on the first sentence typed there: the warm-up run went from 93 slow keystrokes to 0 in Brave. Typing keeps the model loaded. One stutter remains: right after a quick action, if typing resumes within 2 s, about 10 keystrokes, once.

Also in this release

• Quick actions that reword text (improve, professional, friendly, simplify, expand, bullet points, summarize, Fix grammar) keep quoted passages and words written in another script as they are, instead of translating them. A mixture in one script, such as an English noun inside Spanish, is still not solved; listing those words under Terms never to flag keeps them (#1).
• API keys are now also copied to chrome.storage.local, the first step in keeping them out of browser sync. Nothing changes yet for you: keys still sync as before. A later release stops sending them to sync, once every computer has its local copy (PRIVACY.md).

How it is known

On the released tree, npm run verify reported 605 passing checks and no failures. The typing measurements were taken on Windows 11 in Brave 155 and Edge 154 (in-browser model) and Chrome 154 (Gemini Nano, unaffected: 0 slow keystrokes before and after), with the page and runner now in tools/typing-lab/. Translate was also run by hand on Telegram Web for the first time. Measurements and their limits are in MODELS.md and COMPATIBILITY.md.