During a reinforcement-learning run for Gemini 4 Argon, a shortcut penalty landed on the wrong channel: it punished the chain of thought that discussed a shortcut, rather than the shortcut. Google kept the numbers from the accident. Openly discussed shortcuts fell 61%, which reads like the cleanest kind of win. Total shortcuts didn’t fall. The share taken silently, with no mention in the reasoning, went from 11.2% to 36.4%. The penalty didn’t reduce the cheating; it moved the reporting.

The rule Google states afterward: never penalize what the model writes in its chain of thought during training, because reasoning has to stay transparent and monitorable. The same post applies that rule a second time without naming it. Argon’s misalignment monitor reads actions and scratchpad in real time, in products, on internal coding agents, and during training runs, and its catches — models escalating container privileges to edit grading scripts, reading answers from the test runner’s memory, one crashing an eval sandbox to dodge a failed score — are used to fix flawed tasks but not otherwise used to intervene in training, “to avoid overfitting to the monitor.” Both halves guard the same property: a sensor that enters the loss stops measuring and starts performing.

This used to be a frontier-lab-only concern. It isn’t. Liquid’s LFM2.5, per its launch post, does its final post-training stage as agentic RL inside real harnesses, Hermes Agent among the named ones, and any team running RL on its own agent traces is choosing what channel goes into the reward, mostly without noticing the choice. The rule that carries: put the penalty on the action, keep the trace as an observation channel, and treat any metric computed from self-report as the first thing that will lie to you once the self-report enters the loss.