The limits of this measurement
!This isn't the same method as the /ft8-shootout/ article. That was a controlled bench test on WAV files, with different callsigns per file. This is an on-air comparison, on the same antenna and the same audio input: more realistic, less controlled.
!JTDX runs with its hint feature on (Hint=true, Aggressive=5, NDepth=3): in today's session, 56% of its lines carry the hint marker. It isn't blind decoding.
!Search depths aren't matched. Decodium runs adaptive passes at depth 2, 3 and 4; JTDX and WSJT-X stay at NDepth=3; MSHV's configuration couldn't be verified. Each stays as its operator keeps it, not at matched settings.
!Exclusive lines are a clue, not proof. The sender of a line seen by only one program is checked in the other slots: in today's controlled sample it's found 97% of the time. The estimate of ghost decodes remains a minimum, not a maximum.
!MSHV reads about 1 dB higher than the others (median offset +1 dB over more than 3500 messages shared with WSJT-X): the share of «weak» messages isn't directly comparable for it.
!One receiver, one band, one session. Propagation is shared across the four programs, so the differences come from the decoder — but the sample remains a single session, not a run of days.
!Decodium here is 1.0.623, released the same day as this measurement, with fixes to the filters downstream of the decoder: comparisons with earlier sessions aren't homogeneous.
!Open point, not resolved: the same sender appears in two adjacent slots in 0.5% of Decodium's lines, against 0–0.1% for the others — the same before and after the filter fixes, so it isn't a recent regression, but it's still unclear whether these are genuine stations on both sequences or lines built by the deep search.
!Decodium's margin isn't constant — it depends on band conditions. In today's session it ranged from +6.2 messages per slot per second with the band open (12:15 UTC) down to +0.2 with the band closed (14:45 UTC), averaging +2.8 over eleven windows. The comparison only makes sense at a matching point in time: a midday window and a late-afternoon window aren't measuring the same thing.
!The windows from 17:22 to 21:59 UTC on 10/9 are inflated. In that stretch Decodium was running an experimental coherent pass, permanently disabled after the comparison: turning it off dropped the Decodium/JTDX ratio from 1.49 to 1.17, and Decodium's exclusive lines (a sender no other program ever heard) collapsed from 378 and 312 to 0 and 42 across two control windows. These aren't fake as measurements — Decodium really emitted those lines — but they don't represent today's decoder, or tomorrow's.