{"data":{"kind":"file","path":"README.md","version_id":"ua1jivy2aqv305532j0zi45l","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1178,"modified_at":"2026-11-18T05:16:03.424000","content_hash":"a91779f18c1c528b99d810cbbb64df85219a4c5e5e3b32036fa9ad107dbea000"},"entries":[],"content":"# video_conform_repair\n\nRepair a conform that plays cleanly and is wrong.\n\nA conform that plays end to end and is wrong: drifting sync, a ramp at the head, a portrait source forced into a landscape delivery.\n\nEvery fault here is in one class: the artifact runs, raises nothing, delivers,\nand is wrong. Faults that throw are already caught by the software and are not\nworth grading.\n\n## Reward\n\nBinary, 1.0 or 0.0, with a specific reason either way. No partial credit: a\nhandoff is either correct for the department downstream or it is not, and\nrewarding \"nearly right\" teaches the wrong thing efficiently.\n\nShape checks alone are gameable, so every channel or field the repair was not\nmeant to touch is compared against a known-good reference and must match, and\neverything it was allowed to touch must still correlate with the original.\nSatisfying the rules while destroying the data is graded as a failure. On the\nfirst adversarial pass across these environments, seven of twelve attacks\nscored full reward without fixing anything; those holes are closed and the\nattacks are kept in `tools/reward_hack.py` upstream.\n\n## Source\n\nhttps://github.com/AlaboMPJ/alabo-creative-envs\n","encoding":"utf-8","truncated":false,"total_bytes":1178},"status":null}