Enterprises hold decades of operational records. Labs need verified, long-horizon experience to train agents on. Machine57 compiles one into the other.
Every trajectory an agent learns from has to be grounded in evidence, executable in a sandbox, and verified against a contract. Authoring them by hand does not scale. Generating them blindly preserves subtle errors. We spend expert time where the pipeline is uncertain, and nowhere else.
Linked operational events are normalized without erasing their dependencies, then compiled into tasks with a goal, an evidence contract, allowed tools, invariants and a verifier.
Provenance, timestamp and feasibility gates reject what cannot be supported. A reference trajectory is executed and replayed. Nothing enters the corpus on a judgment call alone.
Review minutes go to the tasks with recoverable ambiguity, verifier disagreement and deep dependency chains. Every intervention is logged, replayed and re-checked.
Executable sandboxes compiled from real operational records. Each task ships with its evidence links, tool allow-list, invariants and a state-diff verifier, so it can be used for training and for sealed evaluation.
Trajectories that reached a checked terminal state, with provenance on every action and the review log attached. Sold by domain to labs, or produced from an enterprise's own records for its own agents.
Where should scarce expert time go when building agent training data: authoring every task, reviewing random samples, or repairing selected failures of an automated pipeline? We formulate the expert-hour frontier and study targeted repair against expert authorship and random review under a common budget.
Read the paper (PDF) →Machine57 is a research company. We publish what we find, including the results that are unfavourable to us.
For labs scoping a domain, and for enterprises with records they want to make trainable.