Why "100% of His Cases Were Dismissed" Usually Isn't What It Looks Like

The first time I saw a defense attorney with a 100% dismissal rate in the court data, my gut said what yours probably just did: that can't be real. Nobody wins every case. The gut was right, but not for the reason I assumed. The number wasn't inflated. It was an artifact, the thing data does to you when you take a field at face value.

Here is what sits behind a number like that. In a criminal case, the charge a person is arrested on and the charge they're ultimately convicted of are often not the same charge. Someone is charged with three things, pleads to one, and the other two are dropped. In the record, those two dropped charges read as "dismissed." The conviction lands under a different charge, in a different row, sometimes in a bucket labeled nothing more specific than "Other." Often that plea is a genuine win. Getting the serious charge dropped and pleading to something minor can be the best day of a client's life. That's the craft. But the number doesn't show the craft. It shows the dropped half of the trade and hides the half where the conviction landed, so a hard-fought plea and a real acquittal read as the identical, tidy 100%.

This is why I trust the cleanest numbers in a dataset the least. A rate of 100%, or 0%, or a suspiciously round figure, is far more often a fact about the data than a fact about the world: a missing field, a default value, a category quietly absorbing everything that didn't fit. The number looks authoritative precisely because it's clean. Reality is rarely that tidy, and when it pretends to be, something has been dropped on the way to the page.

I know this because I published the mistake myself. An early version of my own site reported a plea rate for the lower-tier courts, the ones that handle the bulk of cases. It turned out to be an artifact: those courts don't record whether a case ended by plea or by bench trial as a separate field at all, so a default had silently stood in for a fact, across more than two thousand pages. Nobody flagged it. It flattered the story I was telling about the data, which is exactly the kind of number you should distrust most: the one that arrives already agreeing with you. I pulled it, restated it to only the courts where the field actually exists, and left a note in the methodology about why.

So now anything numeric gets re-derived from the source the day it ships, and the derivation saved and dated. Not because it's elegant. Because the alternative is trusting a past version of myself, and that's the person who published the artifact. Correcting a number in public is not a good look, I'm told. I think it's the only one. If you won't show your corrections, you're asking to be trusted. You haven't earned it.

The standard that falls out of this is small and unglamorous. If I can't see how a number was computed, I assume it's an artifact until proven otherwise. If a figure flatters the person publishing it, it gets checked twice. And a number I can't re-derive from the source today doesn't go out under my name, because a statistic you can't reproduce isn't a finding. It's a rumor with a decimal point.

None of this is unique to court records. The same trap sits inside hospital prices, where the "cash price" turns out to be higher than every insurer's negotiated rate for the same procedure in about one comparable line in five in my own data, and a clean comparison between two hospitals usually means someone quietly ignored the ways the two files don't line up. The pattern holds wherever public data gets turned into a number a normal person is supposed to act on: the clean figure is the story people want, and often not the story that's true.

Which is the whole job, as far as I can tell. Not finding the number. Not publishing the number. Distrusting the clean ones long enough to find out what they're hiding.