What layers can a generated intake release contain?

A release for intake can contain several layers, depending on what a lab wants to train or evaluate.

**Records.** Submissions in varied forms (form fields, email text, short call notes), the cases created from them, attachments described as structured metadata, and applicant or account references that can be matched or confused.

**Relationships.** Links between a submission and a case, between duplicate cases, between a case and the policy it falls under, and between a case and the earlier cases it follows.

**Policies.** Written completeness rules, routing tables, escalation thresholds and exception paths, each with an identifier so that decisions can be traced to a rule.

**Tasks.** Prompts such as: decide whether this case is complete, find the duplicate, choose the queue, draft the request for missing information, or explain why an exception needs approval.

**Environment state.** A simulated workspace with queues, a case system and tools, which resets to a known state so that an agent can be run many times from the same starting point.

Not every layer is needed in every release. A buyer can ask for one layer, for example tasks with reference answers, or for the full environment. Scope this in a pilot.

Everything is generated and informed by direct observation of how businesses operate. It is not a set of real business records, and grounding should be inspected in the sample.