Applied AISupport operationsPublic data
Arguments people are certain about, measured before they are trusted.
The subject is variable. The method is constant. Whether the claim is about an AI agent on a support queue, a national grid or a referee, each study writes down beforehand what would settle it, then reports what came back. Three times so far the answer was not the one I expected, and those are the studies most worth reading.
- 6
- studies published
- 3
- findings reversed
- 1
- peer-reviewed
- 1
- recorded talk
Rank 192 clubs on one season of one rate. How much of the spread between them is the club, and how much is the season?
Enough to be worth measuring, not enough to rank on. The same split applies to any table built from one period’s rate. Read the study · persistence.json is where these numbers come from.
Start here
New here? These three are the work reviewed or run in public: one paper, one production study, one talk. Each says what it is and who checked it.
Subjects
A subject earns a place here by having an argument worth settling and public data good enough to settle it. Each is listed because it has published work, not because it is the point; the method is the constant.
Studies
Newest first, across every subject. The label on each card is the claim level the study itself declares, from descriptive to confirmed pattern.
How a claim gets published here
The full contract01
Every number has a script
A figure quoted in a study is produced by a committed script and written to a JSON sidecar published beside it. A number that reached a page without one is how a wrong-signed correlation survived four paragraphs from the sentence that contradicted it.
02
Try to kill the finding first
Specification, aggregation, adjustment coarseness, prior work, baseline sufficiency, and cross-sectional associations across few units. Five attacks and a sixth added after two results died to it. The attempt is reported whether or not it succeeded.
03
Nulls are the output
Claims go viral because they are extreme, and extreme observations are mostly noise. A format that can only ever say look at this is indistinguishable from the thing it set out to correct.
Now
as of
- Running Hagen, the AI agent that drafts the first root-cause analysis for YugabyteDB support, in production. How it works, YFTT 162 →
- Keeping an archive of European electricity data, five countries so far, that refreshes itself every day and re-checks the trailing ninety for restatements. A day on the grid →
- Tracking a pre-registered test of the Portuguese season as it plays out, against a prediction sealed before kick-off. The season as it stands →
Follow the work
New studies are announced on LinkedIn first, with the number and the script behind it in the same post. Everything published here is in the feed, by date.
Text, figures and derived data are CC BY 4.0. Every study links the committed snapshot and script that produced each number, so a quoted figure can be checked rather than trusted.