Threshold Selection

A theta value of 0.99 is described as quality-critical but provides little layer reduction and poor wall-clock behavior in the decision table. A theta value of 0.90 is aggressive and trades higher layer reduction for higher nearest-neighbor failure. The paper recommends theta 0.95 as a default because it balances quality and latency. A stricter training threshold than inference threshold is used to provide headroom under distribution shift. Quality-speed behavior was stable across a range of inference thresholds.