JP Journal · offline evidence lab

Sampling-set inspector

Compare top-k, top-p, and min-p on one bounded probability distribution. Temperature runs first, ties use original index, and one fixed draw is replayed through independently checked masks.

Inputs

Finite logits, comma separatedTemperatureTop-kTop-pMin-pFixed draw [0,1)

Bounded to 2–64 logits. Temperature: 0.05–5. Stable ties: original token index.

Receipt summary

Proof contract

At least one token survives every rule.
Renormalized mass is within 1e-12 of one.
Top-k keeps exactly min(k,N); top-p keeps the smallest reaching prefix; min-p uses a relative threshold.
The draw lands in exactly one interval; repeated canonical input produces the same hash.