{
 "dates": "AI proofread: one session on 2026-10-17 and a rerun on 2026-10-18, 12:57 to 14:56 UTC.",
 "machine": "Mac with Apple M4 Pro, 48 GB memory, macOS 27.0.1",
 "browser": "Google Chrome 155.0.8059.80, 1440x900 viewport, a copy of a signed-in profile. Launched by Playwright with Chrome's sandbox off (Playwright's default --no-sandbox flag) for every run; noticed afterward",
 "network": "networkQuality: 76.3 Mbps down, 66.5 Mbps up, 16 ms idle latency",
 "apps": {
  "word": "Copilot in Word for the web (word.cloud.microsoft), Microsoft 365 business account",
  "docs": "Gemini in Google Docs (docs.google.com), personal Google account",
  "revise": "Revise (revise.io), production"
 },
 "prompt": "Proofread this entire document for any and all issues and mistakes you can find. Leave the text in pristine shape while preserving its original meaning, tone, and wording.",
 "document": "law-of-nations: James Mill, Law of Nations (1825), 16,300 words, 82 errors inserted by ErrataBench, with its answer key",
 "runs": "Five per configuration. Revise (Gemini 4 Argon), Revise (GPT-6 Luna), and Gemini in Docs: the 2026-10-18 rerun, one run of each per round, order rotated each round, no retries; an app's own failure is scored as it stands. Revise (Claude Opus 5.5): added the same day after those rounds, five runs in a row under the same protocol, in Chrome without a visible window (headless). Copilot in Word: three runs from 2026-10-17 and two from 2026-10-18; its other 2026-10-18 runs hit the account's daily AI limit and are in excluded.jsonl. Revise (Gemini 3.8 Flash, GPT-6 Sol, and Claude Haiku 5.5): added the same day after that, five rounds of their own with the order rotated, headless, same protocol; Claude Haiku 5.5 was then left out of the post as a low-cost model (its runs are in excluded.jsonl). Revise (Claude Sonnet 5.5): five runs in a row after that, headless, same protocol.",
 "scoring": "Edits accepted (Copilot's Done, Gemini's Accept all; Revise exports pending suggestions as Word tracked changes, counted as accepted), exported as .docx, and each error checked with ErrataBench's static rule: corrected = expected correction present and error gone; changed_otherwise = error gone but correction absent (not judged); not_addressed = error still present.",
 "timing": "Not reported in the post. Times are in proofread.jsonl where the script measured them exactly (ms = first check that saw the AI's turn over, ms_lo = last check that saw it working).",
 "companion": "Open and import timings: /blog/word-vs-google-docs-vs-revise-speed-test"
}