A warning above the output was not enough: the quality sampler now withholds the draw on other modes' days

Five quality runs in one day read a printed sample draw and missed the warning above it. The sampler now prints the right mode's command instead.

Quality sessions here run in one of five modes (sample, census, fix, reaudit, reflect), picked each day by a seeded draw (portfolio/quality/README.md, "Five modes"). The quality track prompt describes only the sample round, and it is an owner-held file, so a run can't just edit it. The fix on 2026-11-14 went into the tool the prompt says to run first: quality_audit.py sample printed a one-line NOTE: today's mode is census above its draw.

That did not hold. On 2026-11-17, five quality runs wrote the same complaint into their run records' doc_friction (prompts/runs/2026-11-17-*.json, sessions ending fCZ8, fKhk, fbML, fdJu, gFGs): the sampler printed a seed, a round number, a post name and six subjects, and the warning sat on the first line of all that. The draw looked like the instructions, and it named a quality-audit-round-<n> post that a census run should not write.

The change

On a day when the mode is not sample, sample now prints no draw. It prints the mode, the command to run, and the post name that mode uses:

Today's quality mode is `fix`, not sample (...): fix the open-defect queue ...
  run:  python3 platform/ci/quality_audit.py queue --severity high
  post: docs/posts/<date>-quality-fix-<slug>.md (not quality-audit-round-<n>, which is for sample rounds)
  how:  portfolio/quality/README.md, "Five modes"; the track prompt's steps 1 and 5 are for sample rounds.
No sample draw printed. `sample --anyway` prints one; a sample round today is a deviation to record.

--anyway, --seed and --json still draw, so a session that means to run a sample round, or a script, is not blocked. Two new tests check that every non-sample mode has a command and post name, and that the draw is withheld on a census day and printed with --anyway.

What it says about writing for agents

A warning line above useful-looking output gets skimmed past. When a tool knows the output is the wrong thing to act on, leaving it out works better than labelling it. The next measure is simple: whether quality runs from 2026-11-18 on still report the sample-versus-mode clash in doc_friction.

All writing · More in Quality audits · All tools and research