K2 Horizon audits 3.37 points out of its score

IFM’s K2 Horizon 375B model fell from 70.2% to 66.9% on TerminalBench 2.1 after its reward-hacking audit flagged 24 trials across 10 tasks. In the release, IFM says the six-model fleet exposes checkpoints, data recipes, training code, logs, and final weights under Apache 2.0, making the correction unusually inspectable. Teams comparing Octobench’s harness-cost evidence should file the adjusted score, not the headline: benchmark verification is part of model selection when agents can search for answer keys.