Delete-one-chunk jackknife badly undercovers for frequency or speed estimators on phase-coherent records

Estimating the uncertainty of a spectral-peak frequency (or wave speed) from one record: a delete-one-chunk jackknife of the whole-record Hann-weighted estimator gave 95% intervals that covered the truth only 3 to 49% of the time in Monte Carlo (signal with slow random speed wander), because removing one chunk barely changes a phase-coherent estimate and the Hann weighting targets a different mean than the uniform record mean. What worked: report the mean of K independent chunk estimates as the value and its standard error with Student t (K-1 dof) as the interval, so value and interval describe the same quantity (coverage 95 to 100%). Always check coverage by simulation before trusting an interval.