Glossary
Agent
A workflow actor that can plan or choose actions through tools. In production, an agent should be constrained by permissions, state, evals, and human approval gates.
Audit Lifecycle
The business-meaningful lifecycle of a case or decision, such as open, ready for review, approved by human, or rejected by human.
Audit Packet
A durable record containing evidence, versions, transitions, model outputs, human actions, and reasons for a completed workflow decision.
Evidence Packet
A versioned set of documents, extracted fields, citations, and context used by an AI or human during review.
Evaluation
A repeatable measurement of AI system behavior against a task, risk model, and release decision.
Golden Dataset
A curated set of examples with expected and forbidden behavior. It protects important workflow behavior from regression.
Human-in-the-Loop
A control design in which humans own specific workflow transitions or approvals. It is meaningful only when the human can inspect, understand, override, and be accountable.
Idempotency
The property that repeating the same intended operation does not duplicate business effects.
LLM-as-Judge
Using a language model to score or compare outputs against a rubric. Useful for fuzzy qualities, unsafe as the only guard for hard invariants.
Operation Lifecycle
The execution lifecycle of work, such as queued, running, retrying, succeeded, or failed.
Outbox Pattern
A persistence pattern in which state changes and outgoing events are written in the same transaction, then published asynchronously by a separate process.
Prompt Injection
An attack or failure mode in which untrusted content tries to override instructions, exfiltrate data, or cause unauthorized actions.
Semantic Observability
Observability that records AI-specific meaning: task, evidence, prompt version, model version, output, cost, evaluation result, and human correction.
Typed Workflow
A workflow design that represents domain states, events, actors, and transitions with explicit types and validation.