Decision Tracing & Learning Loop
Every approval, override, and edit becomes searchable precedent; the learning loop proposes, humans approve.
Every approval, override, escalation, and edit becomes searchable precedent. The Context Decision Graph is the platform's decision-memory layer; the Learning Loop is what turns that memory into compounding skill.
Context Decision Graph
At commit time — the moment a person approves, overrides, escalates, or edits what an agent proposed — the platform captures the why alongside the what: the rationale, the policy that applied at the time, the approver, and the resulting action. That bundle of signals is the platform's decision-memory layer.
Decision nodes are append-only, immutable, and trace-correlated to the call that produced them. The highest-value nodes are the ones created at moments of human judgment — overrides and exceptions, where senior operators historically left their context only in their heads. Captured as searchable precedent, that judgment becomes shared institutional knowledge.
Learning Loop
The Learning Loop reads completed runs and the decision-memory layer and surfaces patterns worth promoting. Every output is a proposal. Proposals are reviewed and validated before they become active, and a human approves them.
The Learning Loop does not modify model weights. It does not fine-tune. The model is whichever foundation model the team has chosen; the Learning Loop changes only the library of governed skills and the directives that wrap their use.
What this gives operators
"Has anyone approved a case like this before?"
Searchable precedent across the org.
"Why did the agent pick this tool?"
The decision node trace for the call.
"How is the agent getting better?"
A reviewable list of promoted heuristics over time.
What stays out of bounds
Customer data leaving the tenant. Decision nodes and Evidence Cards are org-scoped. Cross-tenant precedent is not exposed.
Autonomous skill promotion. The Learning Loop proposes; humans approve. A proposed change does not affect production output until it is approved.
Model fine-tuning. Changes are recorded in the skill library and directives; they are not pushed back to model weights.
Where to go next
For what operators see day to day, read Agent Operations; for what stays inside your organization, read Learning & Data Boundaries.
Related pages
Looking for the MediaClaw container that produces evidence
Looking for the Media Library this layer reads from
Studying audit, observability, and compliance posture
Studying the discovery model for decision-node surfaces
Last updated