How Reflect Mirror is making decisions.
Every component, every model, every Claude session, every conversation between them — and how those communications turn into wise, researched, evidence-based decisions. This is the nervous system of the project.
Every two minutes, the anomaly_engine reads the live training logs from all 5 GPU machines and flags anything pathological. When it finds something worth killing, the kill_rule_engine picks it up, asks Qwen3 235B Instruct to design a replacement experiment, and drafts the kill+launch plan. That plan lands in your queue — JARVIS never executes destructive operations. You approve from Telegram or the dashboard, and the action moves to Opus or you to run.
Every step is calibration-logged. Every model choice is anchored by a discipline rule (235B-class default, smaller models earn slots only via per-task calibration). Every prompt is versioned and A/B tested before promotion. The whole system is built so that every decision is wise, researched, and evidence-based.