Every message to the yey.boats navigator first hits a small router model. It decides whether to answer instantly from live SignalK data (quick lane) or hand the message to the full assistant (full lane). In production that router alternates between Haiku 4.5 and gpt-5.4-nano (round-robin), and the full lane runs on Haiku 4.5. The custom column replaces both with one local Qwen3.5-4B (yey-router + yey-navigator) on Ollama. Stock Qwen3.5-4B shows what the tuning adds.
Running totals since the demo last restarted, across everyone's questions. "Agrees with live" = same lane (quick or full) as yey.boats production on the same question.
Fields in red are not on this boat (an invented path). The local columns are complete pipelines: they route, then answer either with the yey.boats bundle's real quick-answer code (quick lane) or with that model as the Navigator, with live readings attached (full lane). The live column is a real yey.boats turn, so it is rate-limited. Nothing is executed: no switches, alerts or autopilot. Single-model chat →