Speculative block drafting has an information floor, and current drafters sit well above it.

Block drafters propose several tokens per forward pass, and their rejections mix two different losses that accepted-length metrics can't separate. Defining an information floor, the minimum expected rejection at a given conditioning order, lets you split them. Across four domains, four open-weight targets and a frontier API target, the all-parallel floor reaches 0.286 at the final slot on Qwen3-4B, one realized token removes 86-100% of that floor, and the final-slot model gap accounts for 43-64% of DFlash rejection and 85-92% of DSpark's. Meaning the ceiling isn't the problem yet. The drafters are. (arXiv)