tested on sample size of 100 bills matching the distribution of our data
Compression saves ~11.5% on prompt tokens; completion/reasoning are roughly equal across the two sets.
prompt completion reasoning total
compressed 349,460 157,110 112,128 506,570
uncompressed 394,948 154,710 110,080 549,658
delta −45,488 +2,400 +2,048 −43,088