Leoma Enforces Model Plagiarism Detection, Clarifies Cheating Rules

Leoma introduced a SHA256-based duplicate model check that invalidates miners posting byte-identical weight files, with the earliest committer winning. The team clarified that custom inference stacks and Docker images are encouraged to gain competitive edges, but exploiting the eval pipeline (judge gaming, API outsourcing, metadata faking) triggers blacklist. One byte of weight difference bypasses the duplicate check; the real competition is model quality, not deception.

• •Duplicate model hashes trigger automatic zero score; earliest block wins
• •Custom Docker images with compiled flash-attn and pre-warmed weights encouraged
• •Judge gaming, API outsourcing, sandbox escapes, eval manipulation = blacklist
• •80% task completeness gate real; faster inference = more stable pass_rate