Scrubbed production episodes of a seven-step, tool-using underwriting agent, packaged as RL tasks with ground truth and a programmatic grader.
# grader validation
Episodes 12 (11 complete, 1 incomplete honesty case)
Tasks 12, each with programmatic ground truth
Preference 25 pairs (grounded vs fabricated answer, same history)
Step labels 84, turn level
Grader grader.py, Python standard library only
Agreement 12 / 12 with the human rubric scores
Scrubbed insureds, addresses, phones, emails and signed URLs
replaced with stable tokensNo supplier, broker, client or personal names appear on this page or in its samples. Previews are Perit-owned copies with container metadata stripped.
{"repo_id": "PD-029-r007","primary_language": "Python","lines_of_code": 58214,"commits": 4113,"first_commit": "2016-04-02","last_commit": "2026-08-19","test_files": 212,"ci_configured": true}
The record above is synthetic: the field list and types are the real delivery schema, the values are not from the set. Full field documentation, including every optional column, ships with the pilot batch so you can write the loader before the data lands.
Written spec, rights summary and price inside 48 hours, and a straight answer on anything still in review, including when the answer is that it will not clear.