When should you know the point?

Shreyas Doshi has been thinking about something that keeps surfacing in the work: when a model has done its reasoning and the point is actually visible, vs. when it's still wandering.

The piece frames the problem as a cognitive bias — we tend to treat a model's final answer as the point, even when the reasoning trace has already landed on it and the rest is just padding. The formula the piece points to helps separate the two cases: models that reason briefly and are on the right path, versus models that reason briefly and are being efficient. That distinction is the hinge.

What's useful is the practical test. Full-puzzle generalization — the one that actually rewards what matters — is the real number. If it keeps walking past 27 on regular checkpoints, the trajectory's confirming. If it sits at or below 27, the peak you celebrated was a peak-pick, not a peak-real. The mean-nets and per-net scores are worth noting but they're the reward-adjacent family; the full-puzzle score is the one that belongs in the pitch.

The piece also flags something subtle: behavioral shifts in the policy often show up before the per-token entropy readings catch on. A 0.0093 nat reading can mask a directional shift in mass that's already nonzero. Esto te toca — the model is already changing, even if the metrics haven't caught up.

Why this matters for us: the same question applies to our work — when has the insight landed, and when is the rest just noise? The formula gives us a way to tell.