The evidence layer for the execution environment. Not what the model processed, but the conditions under which it operated.
Structured signals that describe the execution environment: permissions, memory state, tool availability, retrieved knowledge, agent state, and runtime conditions at the moment of the AI call.
They are not the input prompt and not the output. They are the surrounding evidence that explains how the system arrived at the input it sent to the model.
Captured by the SDK at the AI call boundary, normalized into the canonical schema, and bound into the CER alongside inputs and outputs. When included, they are protected by the same SHA-256 certificate hash.
Context is optional. When present, it is part of the integrity guarantee of the certified record. When absent, the record still verifies on its inputs and outputs alone.
Auditors can inspect the conditions of a decision, not just the result. Reviewers can see what the system knew when it acted.
For multi-step agents, context signals at each step produce a verifiable trace of how the agent reasoned, retrieved, and acted across the run.