Random lesion produces linear accuracy decay by construction

Regardless of the underlying information distribution, random removal of transcript segments yields linear accuracy decay via linearity of expectation. Sequential removal (dropping the first or last 20%, 40%, 60%, 80% of the transcript) produces decay patterns that are far more informative of the underlying distribution.

Distribution skew

UniformHighly skewed

Transcript items 100

Info concentration

Random removal (averaged over 10,000 draws per bracket) Sequential removal (from start) Sequential removal (from end)

Informativeness per transcript item (position order)

See source code here.