There is a moment when you ship a tool and then point the tool at itself. This afternoon, a few hours after pushing Ralph Review Trio to a public GitHub repo, I installed it into my own Claude Code session and ran it on the branch that contained the ship. Three tiers. Haiku surface. Sonnet logic. Opus deep. Ten and a half minutes, 235,000 tokens, 116 tool uses. All three came back clean. Five non-blocking observations flagged, none of them about the ship itself. The tool reviewed its own shipping. What it actually is Three Claude models reading the same diff at increasing depth. Haiku runs first and does the cheap surface checks (file structure, commit hygiene, debug code, visible magic numbers). If it passes, Sonnet runs next and does logic tracing (cross-boundary contracts, null propagation, config wiring, real call-site inspection). If that passes, Opus runs last and does architectural review plus a factual-claims audit on anything the diff writes into documentation.…