Every object has a contract

An autonomous research system produces artifacts faster than a human can audit them. The only way that scales is if the artifacts audit themselves — if a malformed environment, an ungrounded claim, or a silently redefined reward cannot pass through the system unnoticed.

Our answer is an ontology: seventeen primitives that cover the vocabulary of a reinforcement learning workspace — environment, task, observation space, action space, learning signal, evaluator, run, and the rest — each with a versioned JSON Schema, bindings, and conformance tests.

Validation at the boundary

Contracts are only as good as their enforcement points. Records validate at every operation boundary: when an agent writes an object, when a run consumes one, when a value crosses an edge in an orchestration graph. Typed edges check kind, cardinality, and schema before a value transfers. A workspace that passes validation is a workspace that runs.

Closed-graph validation goes further: given a bundle of records, it checks that every reference resolves, every binding matches, and nothing dangles. A result is not just a number — it is a graph you can walk from the claim back to the exact definitions that produced it.

Why this matters for training

The point is not tidiness. The point is that reinforcement learning is uniquely sensitive to specification drift. A reward that shifts meaning mid-experiment, an observation space that silently widens, an evaluator that diverges from the deployed one — each produces results that look right and are not.

Contracts make drift a type error. When every experiment snapshots the exact versioned definitions it ran, a result you read six months from now is a result you can re-execute. That property — reproducibility as a side effect of the data model, not a discipline you hope researchers maintain — is what lets the lineage compound.