Live MCP session — payloads captured 2026-04-03 from Graphonomous v0.3.1

A/B: Q-Value Retrieval — Before vs After Outcomes

Same retrieve_context query, two phases: A retrieves with similarity-only ranking (no outcome history). B retrieves after outcomes have been recorded — Q-values shift the ranking so high-confidence-but-low-utility nodes drop below high-utility ones. Two-phase retrieval is self-correcting.

Chat + tool-call storytelling
Live tool-call log
idle