This red-team run seeded 24 small faults, one at a time, into three files, and ran each file's own tests after every fault. 14 faults broke a test, a catch rate of 61%. One of the ten that got through could not change anything (see below). The run added tests for the other nine, and the rerun confirmed the new tests catch the five faults it can measure on its own. That puts the rate at 83% by the tool's count, and every fault is now caught once CI's shared contract runs too.
A bid calculator that assumed zero
The basement finishing bid calculator reads about forty numbers, and every one falls back to 0 when the caller leaves it out (input.hoursPerBedroom ?? 0). Two faults changed a fallback to 1. No test noticed, because every test passes every field. A shared link or a script calling the library can leave a field out, and then the quote would include an hour per bedroom that nobody entered.
The new test covers all forty fallbacks at once, not just those two. For each numeric field in two sample jobs, leaving the field out must give the same quote as passing 0.
A day stamp with a trailing T
Most of the portfolio's 1,286 tools carry their own copy of a small on-device usage counter. It keeps counts in the visitor's own browser, never sends them anywhere, and stamps the day as new Date().toISOString().slice(0, 10). One fault turned the 10 into 11, so the stamp read 2026-11-18T. Nothing caught it: not the tool's suite, and not the shared contract (platform/ci/usage_contract.mjs) that CI runs on every copy. Neither one ever read the stamp's value.
The contract now checks it. A field with "day" in its name, or a key shaped like a date, must be a bare YYYY-MM-DD for today, by UTC or local time. The first run of the new check passed all 1,286 copies. It rejects the .slice(0, 11) fault and a full timestamp put where a day belongs. Both are now test cases in platform/ci/test_check_storage_guard.py.
The counter's other three escapes were already caught by that contract. The red-team tool measures only the tool's own suite, so it still counts them as escaped. That gap is logged (2026-11-15) and needs a change to a control-plane file.
The deploy page's command line
The third target was platform/deploy/build.mjs, which writes the one-click deploy page. Three faults in its command-line block got through: reading the wrong argument, flipping the "am I the main module" check, and cutting the --out= value one character short. The tests call build() directly, so they never run the command the pages workflow runs. A new test runs node build.mjs --out=<temp dir> and checks the output folder and the summary line. The fourth fault was harmless. It made one mkdir call non-recursive, but the line before it had already created the parent folder.
One note for the next run
A target whose tests import numpy gives a "red baseline" under the container's bare python3. That was the first draw here (platform/audio/test_narrate.py). The fix is to run redteam.py with the interpreter that platform/ci/pyenv.py pytest numpy prints.