| Candidate | Domain | Source | Turns | Discrim. | Verdict |
|---|---|---|---|---|---|
| reconcile-q3-ledger | spreadsheet ops | claw-gdrive | 31 | 0.84 | promote |
| redline-msa-indemnity | legal drafting | trace/8821 | 44 | 0.81 | promote |
| parcel-join-tolerance | GIS | trace/7710 | 19 | 0.77 | promote |
| refold-chain-b | bioinformatics | pdb-refold | 26 | 0.71 | review |
| icd-ambiguous-dx | clinical coding | trace/9042 | 12 | 0.66 | review |
| cold-chain-reroute | logistics | trace/6633 | 38 | 0.59 | review |
| calendar-dst-edge | scheduling | claw-gcal | 9 | 0.31 | too easy |
| slack-thread-summary | support ops | claw-slack | 7 | 0.18 | not verifiable |
Oracle solves it, but two of six models pass by guessing the seed. Verifier probably needs a tolerance band.
Two defensible codings. Needs a rubric before it can be scored automatically.
Long horizon, high variance. Candidate for a multi-scene split.