How do schedule workflow trajectories fit into a model lab's existing pipelines?

Schedule material can be delivered in shapes that match common lab pipelines, and the format is something to agree in scoping.

**Trajectory form.** Each episode is a sequence of turns: an initial state snapshot, the policy text in force, a request, the agent's tool calls, and the final state. In JSONL, one line per episode keeps splits and sampling simple.

**Typical top-level fields:**

- `episode_id`, `domain`, `variant_seed` - `initial_state` with `accounts`, `schedules`, `instalments`, `mandates`, `approvals` - `policy_refs` pointing to versioned policy documents - `turns` with role, message and `tool_calls` - `expected_final_state` and `grader_spec`

**Uses.** For supervised fine-tuning, the turns become targets. For preference data, a pack can pair a correct trajectory with a plausible faulty one, such as an unapproved extension. For evaluation, the harness loads `initial_state`, lets the model act, then diffs the result against `expected_final_state`.

**Tables.** Where a lab prefers analytics, the same content can be flattened to Parquet with keys such as `schedule_id` and `instalment_id` linking the tables.

All content is generated. Ask for a sample in your preferred format and check that identifiers resolve across every table before building a loader around it.