Skip to content
MirrorFirm

What technical handoff should a lab specify before starting?

A technical handoff can define formats, schema versions and reproducible reset behaviour so a lab can run generated tasks inside its own evaluation stack.

Agree a release contract that matches the lab's ingestion and execution systems. JSONL can carry task episodes, while Parquet can carry linked record tables. Specify encoding, timestamp conventions, nullable fields, attachment references and schema versions rather than relying on format names alone.

A manifest can include file checksums, policy versions, environment configuration and the mapping between task identifiers and starting states. Keep evaluator-only answers outside the model-visible payload, with an explicit loader boundary to prevent accidental answer exposure.

For simulated environments, request a documented reset interface that accepts an episode identifier and seed. State which elements the seed controls and whether tool responses remain deterministic. A reset should return a state fingerprint so the lab can verify that repeated runs began from the same configuration.

Before expanding scope, ingest one complete episode and execute its reset, observation, action and scoring cycle. This first technical step tests compatibility without presuming a particular hosting arrangement. Record any required adapters as part of the agreed handoff so later results remain attributable to the same execution contract.

Resolve packaging and execution assumptions early enough to make the pilot reproducible.

Audience:
Model labs
Content type:
Integration
Buyer question:
Scope and getting started

Content release: r-a220eb6e099b243b

More on Overdue follow up