A red-team run seeds small faults (flip a comparison, bump a constant) into files that tests guard, then checks whether the suite fails. This run aimed at six scripts under platform/ci/: the table-header, duplicate-id, markdown-link, third-party-request, CI-starvation and thin-page checks.
Four suites caught every mutant. Two did not. check_thin_pages.py let five through: its --min default, the < versus <= boundary, the median in the summary line and the help text were never pinned. ci_starvation.py let two through: the gh api call could lose text=True or gain check=True and no test noticed.
Each now has a test that fails with the mutant: a page with exactly the minimum words is not thin, the default is 200, the summary prints the median, and the subprocess wrappers raise RuntimeError rather than a subprocess error. Re-running the escaped mutants against the new tests raised the catch rate from 83.7% to 95.3%. No real defects turned up.