The repository audits its own finished work against a senior professional's bar, and a "fix" session then works the open-defect queue that the audits leave behind (python3 platform/ci/quality_audit.py queue). The queue listed defects worst first, then oldest audit first. So did every other session's copy of the queue.
On 2026-11-16 two fix sessions started within the same minute. Both ran queue, both took the first medium-severity row whose subject had no recent commits (research/sequential-slashing), and both began fixing it. The queue already marks a row once an open pull request names its subject, but neither session had opened a PR yet, so neither could see the other. One run record reported it in doc_friction.
The sample and census commands hit the same race earlier and solved it the same way: a random tail on the day's seed, so concurrent sessions draw different subjects. queue now does that too:
- Within a severity, subjects come in a seeded shuffle. A subject's defects stay together and in order, and high-severity rows still lead.
- The seed is printed, so
--seed <s>repeats an order;--oldest-firstgives the old fixed one. - Rows an open PR already names drop to the bottom of their severity, so the top row is one nobody has claimed.
With 391 open medium-or-worse defects across many subjects, two sessions now start on the same subject only by chance. A new test checks that different seeds lead with different subjects, that a seed repeats, and that severity order holds.
The same run fixed a smaller snag reported twice on the same day: quality_mutate.py, which seeds a bug in a private copy of a subject and runs its tests, ran the test command without a shell. A command such as PYTHONPATH=src python3 -m pytest failed because the system took PYTHONPATH=src for the program name. A command led by VAR=value now runs through bash.