Part 1 of 4 · 1 min read
Overview
This level's project criteria ("a rollback has been tested") are about this module. Agent releases fail in new ways: a prompt edit can break behaviour without a code diff, a secret in an environment variable can leak through a trace, and a long-running run can straddle a deploy. Interviewers ask "how do you ship a prompt change safely?"
About the example, and what is real
Real: the scanner, Dockerfile lint, significance-based gate, rollout simulation, registry (tags versus digests), migration compatibility check and the 22 test cases in code/mini-deploy. Assumed/invented: the 16 fixture lines, 20 percent regression rate, 5-point regression, 80 percent baseline, eval-set sizes, the bad release's 20 percent error rate, 8-minute human detection, 1-minute canary detection, 50 requests per second. Not real: no CI service, container build, image scan, orchestrator, database or real traffic; the lint and scanner are toy versions of real tools.