Whitepaper
Internal study: replication of the Token Saver benchmark on a Claude Code host via the /api/mcp/claude mount - 28.9% fewer visible tokens and 83.4% lower cost at held quality (14/15 vs 15/15 rubric checks), ~18s planning latency, end-to-end ledger settlement, plus automation-planning findings, the verified settlement/decomposition fixes, and the three-arm quality-floor experiment the current evidence still requires.