How should a pilot delivery be structured for easy ingestion?

A pilot delivery is easiest to ingest when it behaves like a small software release.

**Package layout** - `manifest.json` with release ID, schema version, seed, file hashes and split assignments. - `schemas/` holding JSON Schema for each table. - `data/` with JSONL for cases, events and tool traces, and Parquet for larger tables. - `policy/` with the authority matrix and policy documents in both machine-readable and readable form. - `expected/` with outcomes, held apart so they can be excluded from agent context. - `graders/` with check functions, a rubric file and example outputs.

**Environment interface** - Tool actions such as `read_case`, `read_policy`, `list_exposure`, `refer_case`, `approve_case`, `decline_case` and `return_for_info`, each described with typed arguments. - `reset(seed)` restoring initial state, and `snapshot()` returning current state for grading.

**Practical points** - Pin the release ID in your harness config so results stay comparable when a later version appears. - Run a smoke test on the sample: load, validate against schemas, reset, execute one scripted trace and grade it. - Agree who owns issues found in the sample, and how corrected versions are issued.

Delivery arrangements and any hosted environment access are agreed in the engagement, so confirm them before planning integration work.