Backup

Entropy-based confidence gate for decision-making

Ht = − Σi pi,t log2 pi,t

Signal recovery of the top hypothesis when

H̄t < H([0.9, 0.1]) ≈ 0.47 bit

H̄t: moving average of entropy over the last evaluations. Threshold = one hypothesis holding 90 % of the mass against one rival.

entropy time threshold 0.47 spike act raw H smoothed H̄

1 · Evidence wait

≥ 2 detected parameters → window of 5 evaluations. One parameter → longer window (10).

2 · Smoothing

Average entropy over the window, so a one-tick dip cannot trigger action.

3 · Hold

Never act when “No Anomalies Present” leads; restart the window.

input: any normalized score vectoroutput: recover(target) or waitno per-method tuning
habitat-dx · multi-agent AFDDR simulation