Mint team experiment note, September 26, 2026
Published excerpt: October 15, 2026
Wording edited October 16, 2026: internal identifiers and implementation details
left out. Every measurement below is unchanged.

- Copies of four real Codex conversation files (never the originals; copied with
  the reading process's access-date updates off, and their access dates checked
  unchanged), compressed by a development build of Mint with a file-level
  compression method: 4 of 4 compressed, 3.13 GB in 12.6 s end to end (compress,
  file identity kept, every byte read back), blocks 0.445, 0.606, 0.624 and 0.627
  of before, reported saving 1,285,251,072 B = the measured block delta, `df`
  +1,255,124 KB, estimate shown before the run 1,163,402,034 B, every copy
  byte-identical by sha256. Codex's own app-server reopened them exactly as uncompressed copies:
  3 threads (the 564 MB and 1.15 GB files are the second and third files of one
  thread, reopened with its 44 MB first file), 1,598 turns, 2,381 items, and the
  original content stayed intact after a resume.