Each research programme here has a README that a reader opens first, and a LOG.md that keeps the full record. In two programmes the README had slowly turned into a second log, one sentence per session.
What we found. The build-data README, the programme asking whether open permit records predict permit and build times better than each city's own averages, had its whole results section on one line: 21,344 characters, milestones 1 to 7 plus twenty follow-up sessions (77 to 96) on the build-time transfer gap. A run record this morning said finding where to add a result "takes a scripted search". The llm-coding-inference README held one 12,857-character paragraph covering milestones 7a to 7aw. A scan of every research/*/README.md and ventures/*/README.md found no other paragraph over 6,000 characters. The next longest in research is 4,279.
What we changed.
build-data's README now has a short status, a "Results by milestone" section, a "Milestone 6 follow-ups" list with one bullet per session, and a "Reproduce" section. Thellm-coding-inferenceparagraph is split at its own bold run-in headings into 17 paragraphs.- No words changed. A script compared both files with whitespace collapsed before and after and found them identical, apart from the three headings and one line asking new sessions to add a bullet.
research/_shared/programme_check.py, which research CI already runs, now fails a README line over 6,000 characters. It names the line, the length and the fix (sections, or one bullet per session, with the detail inLOG.md). We checked that it flags the oldbuild-dataREADME (line 3, 21,344 characters) and passes every README onmainnow.
Why a check and not a guideline. The paragraph grew because each session did the locally sensible thing: it added its result next to the last one. A sentence asking people to write shorter paragraphs would need every session to read it. A check that fails only when a paragraph actually gets too long costs nothing until then.