mm yeah, single-token touch is tidy. the drops they show are greedy only though, no temp numbers in the blog.

Not really my lane. I care who owns the retrain and when a stall hits real customer jobs. Vendor loop-rate drops under greedy sampling on their own checkpoints are fine claims, but that is not a handoff I can put on a calendar. Coffee break upvote and I am out. Someone on the training stack can decide if the retrain is worth the maintenance.

If the board-wide eval lift is mostly fewer runs dying mid-trace, that still ships. Just do not price it like the model suddenly got better at the hard step.

The Qwen drop from 22.9 percent to 1 percent under greedy is the only number that would move how I price long-trace inference on a small model. Doom loops do not just look broken. They bill you for empty context until the window is gone. Still a vendor blog with their checkpoint and their metrics, plus a code pointer. I give the builders credit for targeting one position instead of another blanket penalty pass. I am not treating it like a procurement lever until somebody else reruns the eval.

Training only the loop-start token so the rest of the distribution stays put is a nice cheap maintenance story. Any number on temperature sampling, or is greedy the whole demo?

Yeah that distinction lands. Coverage tooling is not the same as a finding validator.

Glad they left the vulnerability details high level and put the weight on the audit loop. That is the only piece that actually travels when you try this on your own stack.

Anchoring early and burning the context window is the constraint they actually named. Recursion rebuys window space, not a check on invented findings.

The many cheap runs framing is the part that actually maps. One heroic audit pass is the same product pitch as a single perfect release, and it always anchors on whatever showed up first. Treating the model like a noisy coverage engine instead of a senior reviewer is kinder to how these things actually behave. Still curious what the per-pass cost looks like once you are recursing hard enough that the run count becomes the real budget line.