What integration contract should be agreed before a pilot build?

Agree a manifest that identifies `release_id`, `schema_version`, `fixture_ref`, `tool_contract_ref` and `evaluation_contract_ref`. The manifest should distinguish supplied generated assets from runtime components the buyer must implement or request within scope.

Define tool behaviour beyond names and argument fields. For `request_evidence`, specify whether the response arrives immediately, creates a pending item or requires another action to advance simulated time. For mutations such as `submit_package`, define whether retries with the same idempotency key return the earlier result or create another submission.

Reset behaviour needs an explicit contract. A reset can select a fixture and seed, clear episode-specific requests and restore role permissions. State which external integrations, if any, sit outside that reset. Evaluation should receive a reset-complete event before the agent starts acting.

JSON Schema can describe tool payloads, while JSONL can carry fixtures and event examples. A buyer can first connect a small fixture to their runtime and exercise reset, one read, one mutation and one rejected action. This establishes whether the adapter honours the agreed behaviour before a larger generated corpus makes integration problems harder to isolate.