DMSD Phase 3G — Structured Null Escalation
逐步增加 divisibility source depth，觀察 underdispersion 與 residual lag-1 在哪裡消失。
Decimal scale
Primary residual model
1000 bins · candidate counts after trial division ≤ y · smooth 1/log x weighting.
Closure threshold
First lag sign flip
Best variance match
Min RMS threshold
| y | precision | lag-1 | RMS | expected RMS | ratio | max p | state |
|---|
當 max conditional p > 1 時，smooth-weighted candidate-binomial baseline 已失去 probabilistic admissibility；靠近 trial-division closure 時，source 本身已吸收太多 density structure，不能再盲目疊加原來的 1/log x correction。