Human-Agent Assurance Framework
Version 1.0 - Snowflake workflow control model
Control windows
- Pre-action: policy checks, role validation, and declared scope.
- In-action: trace prompts, tool calls, model outputs, and approvals.
- Post-action: validate outcomes, review overrides, and seal evidence pack.
Minimum evidence set
- Human request and agent plan trace.
- Execution timeline with identity, role, and authority context.
- Approval and override events with rationale.
- Final outcome and verification checks.
Causal discipline
Post hoc does not imply propter hoc. Causal claims are accepted only when measurable, repeatable, and independently verifiable under the same control conditions.
Snowflake-relevant examples
- Cortex-assisted SQL refinement with human approval checkpoints.
- Agent-recommended model or pipeline changes with post-run evidence verification.