train an agent to make high-accuracy policy decisions
claims, triage, moderation: an agent trained on your reviewers’ past decisions can reason through a case, decide it accurately, and write down why.
a claim comes in. deciding it is not one classification. does the policy cover the event, which clause governs, is the amount inside the band, do the documents corroborate the story, is there a fraud signal, pay it or deny it or send it up. then the part that keeps the whole process defensible: a written justification that cites the policy. moderation queues, support triage, and kyc reviews all have this shape. several interdependent calls per case, made under written rules, leaving a paper trail.
why a classifier never quite fit
a classifier returns a label and a confidence. it can’t weigh two clauses that pull in opposite directions, produce the six related decisions a case actually requires, or write the paragraph your auditor and your customer will read. so the pipeline stays human everywhere it matters, or automates a thin trivial tier while the backlog accumulates on the rest.
your case history is a curriculum
the training data is your archive: the case as it arrived, the decision your expert made, and often the note explaining why, next to the policy documents themselves. supervised finetuning on those decisions teaches a model the shape of the work; we’ve shown a 4b model beating an untrained frontier model on a structured extraction task after 3,000 examples. reinforcement learning then sharpens it against decision accuracy, scoring the whole bundle at once, so the model is rewarded for getting the labels, the clause citations, and the justification right together, not for one lucky field.
reasoning you can read, escalation you can tune
the output is the full decision artifact: the calls, the extracted fields, the cited clauses, and a narrative justification in your house style. below whatever confidence bar you set, the agent escalates to a human instead of guessing. precision on the automated tier is measured against held-out historical cases before anything goes live, and the escalations your reviewers keep handling become exactly the training data that improves the next run.
how it runs on castform
export historical decisions and the policies they were made under. we shape them into training data and evals, run sft and rl on managed infrastructure, and report precision, recall, and calibration at the thresholds you care about. the weights are yours. and when the policy changes, you retrain on the new rules instead of stacking another paragraph onto a prompt.