• What: Given a small argument graph (nodes/edges), OpenLine returns a verifiable receipt (claim/because/but/so + guardrails + telemetry).
• Why: Tracing is descriptive; receipts are normative. They make agent steps auditable & safe by default.
• How: stdlib Python; MCP JSON-RPC stub; LangGraph node; schema + validator; optional signing step; GitHub Actions + Pages.
Repo:
https://github.com/terryncew/openline-core
Demo:
https://terryncew.github.io/COLE-Coherence-Layer-Engine-/nca...
Would love feedback on the guard checks and the OLP schema—where should it grow?