Across 33 models from five families, fixed quantization error accumulated only when the model’s loop failed to settle. Settling models instead converged to a shifted state, so running them longer did not keep worsening the answer.
The shift can be compared with the model’s tolerance using a label-free sensitivity measurement. Failed models can also be recovered with a few final 8-bit loops. A halting-based controller using this strategy beat fixed-depth inference by up to 15 points on Sudoku-Extreme and Maze-Hard while using under a third of the weight traffic.