Scripts and raw outputs behind aibuilderclub.com/blog/jev-engineering-guide (run 2026-09-22).

jev-guard.mjs      PreToolUse hook (Claude Code). guard-run.txt = first run (before the fail-closed fix); guard-run-2.txt = the run shown in the article.
route.mjs          model/effort router. route-run.txt = the four tasks, verbatim, with usage.input_tokens and ms.
log-cron.mjs       the log-triage cron. run-1.txt = the 600-entry replay. rubric.before.json = rubric at the start of the run; rubric.json = after Claude Code's patch.
                   Run it as: LOG_PATH=/path/to/your.log node log-cron.mjs   (needs TYPESAFE_API_KEY, or a proxy that injects it, and the claude CLI on PATH).

Two lines were tightened after the run and are marked in the article: the hook now rejects probabilities outside [0,1], and the cron rejects a patch that names an existing class. page_on_severity was lowered from 2 to 1.5 (a weighted mean on a 0 to 2 scale never exceeds 2). None of the three changes the outputs recorded here.