Canonset
← Samples

RL environment · business software

Expense approvals

A practice copy of Alder & Vale Freight's expense-review app (a fictional company), a written policy, and 50 tasks an agent is scored on automatically. Every task is proven solvable, and proven not to reward blanket strategies.

Verification, recomputed on every page load

Share of checks each strategy passes. A task is valid when the reference scores 100% and no attack reaches 100%.

TaskReferenceDo nothingApprove everythingReject everythingAsk for receipts everywhereValid
exp-01100%60%40%0%0%valid
exp-02100%60%0%0%40%valid
exp-03100%43%14%14%0%valid
exp-04100%50%17%0%0%valid
exp-05100%50%50%0%0%valid
exp-06100%50%17%0%0%valid
exp-07100%50%50%0%0%valid
exp-08100%43%14%14%0%valid
exp-09100%43%14%14%0%valid
exp-10100%43%14%14%0%valid
exp-11100%60%40%0%0%valid
exp-12100%60%0%0%20%valid
exp-13100%60%40%0%0%valid
exp-14100%50%17%0%0%valid
exp-15100%60%0%0%20%valid
exp-16100%50%17%0%0%valid
exp-17100%60%0%0%0%valid
exp-18100%60%40%0%0%valid
exp-19100%50%17%0%0%valid
exp-20100%60%0%0%0%valid
exp-21100%43%14%14%0%valid
exp-22100%43%14%14%0%valid
exp-23100%57%43%14%14%valid
exp-24100%50%0%0%17%valid
exp-25100%60%0%0%20%valid
exp-26100%38%13%13%0%valid
exp-27100%67%0%0%17%valid
exp-28100%60%0%0%0%valid
exp-29100%60%0%0%0%valid
exp-30100%60%0%0%40%valid
exp-31100%60%40%0%0%valid
exp-32100%50%17%0%0%valid
exp-33100%50%17%0%0%valid
exp-34100%50%17%0%0%valid
exp-35100%38%25%13%0%valid
exp-36100%44%33%0%0%valid
exp-37100%0%42%8%25%valid
exp-38100%33%33%11%0%valid
exp-39100%43%29%0%0%valid
exp-40100%30%30%0%10%valid
exp-41100%27%18%9%0%valid
exp-42100%43%0%0%29%valid
exp-43100%60%0%20%0%valid
exp-44100%67%0%0%50%valid
exp-45100%60%0%0%0%valid
exp-46100%43%14%0%0%valid
exp-47100%43%14%0%0%valid
exp-48100%27%9%27%0%valid
exp-49100%20%30%10%10%valid
exp-50100%0%50%0%17%valid