A quality-audit session in this repository reads two instructions in a row. The track prompt says step 1 is quality_audit.py sample, which draws six subjects for a fresh audit round. The area README, one link down, says a session runs one of four modes (sample, census, fix, reflect) and that quality_audit.py mode picks which, by a draw seeded on the date that doubles the weight of fix while high-severity defects are open.
Both are right about their own layer, and they disagree about what a session does first. platform/ci/friction_rank.py --live counted ten separate quality runs between 2026-11-10 and 2026-11-14 that reported it in their run records: "Step 1 says run sample, but the README says mode picks among four modes first", and "the prompt's step 5 asks for a round post that fix mode doesn't produce". Each one stopped to work out which to follow.
The obvious fix is one sentence in the prompt. The prompt is owner-held, though (it sits under .github/CODEOWNERS, so an edit waits for approval), and a sentence every quality run must read is a cost paid every run. The cheaper fix sits in the command the prompt already sends people to. Today sample checks the mode before it draws:
NOTE: today's mode is `fix`, not sample (... 5 high defects open): fix the
open-defect queue (queue), with a regression test per fix, then re-audit.
portfolio/quality/README.md, "Four modes", says how; a sample round today
is a deviation to record.
seed 2026-11-14-a8c5 round 74 ...
Today's draw is fix, so a run that follows the prompt literally now sees that before anything else. When the mode is sample, or a seed is passed by hand, the line does not print. The README's mode table gains two lines saying the round post belongs to a sample round and that a fix session names its post for the fix, as this morning's quality-fix posts already do.
The measure for whether it helped is the same count: quality-run friction reports about the prompt and README disagreeing, from 10 in the five days to 2026-11-14 down to none in the five days after.