Input and context
Identify who supplied the material and which application or retrieval boundary admitted it.
Follow a proposed action from model output to the system that can execute it. Ask which identity is acting, what policy permits and which record shows the downstream result.
Open the interactive architectureFor the complete component catalogue and its video walkthroughs, see the AI Harness capability matrix.
A model can call a tool directly, and the record shows only the resulting action.
A scoped identity and policy gate authorize the tool call. The record joins the decision, approved scope and actual outcome.
YAML and JSON show release settings. Python illustrates the reviewer-role and approval checks. The application must obtain those inputs from trusted services, bind approval to the action and record, enforce the decision at execution, and record both the decision and outcome.
action: release_record
allowed_role: quality_reviewer
require_approval: true
record_decision: true{
"action": "release_record",
"allowed_role": "quality_reviewer",
"require_approval": true,
"record_decision": true
}def permits_release(role, approved):
return role == "quality_reviewer" and approvedDiscuss an AI Harness assessment
Find definitions, affected functions and defensive evidence to review.
Compare endpoint signals with the AI Harness evidence you need to verify authority, runtime decisions and containment.