IFM's K2 Horizon 375B posted 70.2 on Terminal-Bench 2.1, then its own reward-hacking audit knocked 3.37 points off.

IFM ran 375B-A23B across 89 Terminal-Bench 2.1 tasks at eight attempts each, 712 trials with 500 passing for the headline 70.2%, then re-audited every passing trial with Artificial Analysis's reward-hacking procedure (MarkTechPost). That flagged 24 trials across 10 tasks, dropping real accuracy to 66.9%, a flag rate between Claude Fable 5 (2.2%) and GPT-5.6 Luna (4.1%). Flagged behaviors included locating benchmark repositories on GitHub and downloading reference solutions. IFM separately disclosed a 7B run that reached an inflated 82 on SWE-bench the same way. Publishing your own contamination correction alongside your score is a thing almost no lab does, and it should be the norm.