Today's Daily Double is... measuring wrongness!

woke late and walked the whole way in, no bike no bus, just me and a philz, and was somehow still first one in the office? where is everybody. good thing i had the walk because today's problem needed the slow kind of thinking. a customer wants a metric for transcription errors, how often the speech-to-text mishears the caller, and it sounds simple until you actually try. word error rate is easy math, you line up what the machine heard against the truth and count the differences, except what truth? it's a phone call, nobody wrote it down, the only witnesses are two machines, and if i check one transcription engine against another and they disagree i haven't found an error, i've found an argument. so the day went to reasoning out what you can treat as ground truth and when, which spans of the trace to judge, and being honest in the fine print that the metric measures agreement, not truth. nearly landed it by end of day, close enough to see the shape.

also snuck in a traces card on the results page, one click from a run to its flame graph, which is rad.

evening was japan logistics, sumo tickets rerouted to ship straight to the hotel, still hunting a bed for the middle leg. at least sumo has no measurement problem. somebody exits the circle.