Closing the Loop: Empirical Model Routing Calibration from Production Observability Data

Multi-agent LLM systems deploy model routing — the assignment of operations to specific model tiers (e.g., Sonnet, Haiku, Opus) — based on intuition or static rules established at deployment time. These rules are never systematically updated from production evidence. This paper presents a closed governance loop that connects production session observability, quasi-experimental impact analysis, operation criticality classification, model routing, and economic circuit breaking into a continuously self-calibrating architecture. The synthesis draws on five companion works in the Gallora Lab series and introduces a calibration algorithm grounded in empirical production data (session of 2026-04-03: 278,730 output tokens, $36.74 total cost, 94.2% cache hit rate, 19 artifacts produced). The central empirical finding is that cache architecture dominates cost with 9× leverage while model tier selection has approximately 1.1× leverage on output tokens — which represent only 0.43% of effective token volume. The governance implication is counterintuitive: optimize cache architecture first, then calibrate model routing from quality evidence rather than cost assumptions. Four reclassification triggers, a conservative minimum-evidence protocol (N ≥ 30 before any downgrade), and six interface specifications are defined for implementing the loop in production.