Short reports from the gap between plausible and works. Fortnightly, under 500 words, and every claim carries a receipt. We change our methods in public; this is where we report what worked, and what didn't.
Fortnightly. Under 500 words. Unsubscribe is one click and no hard feelings. Privacy.
We ran an experiment on ourselves. A fresh AI, no context, no access, no memory of us, was given everything a visitor to our website sees: the page copy, the fourteen free notes we publish in full, and the concept map, all fifty-four titles and how they connect. Its job: rebuild our licensed OKR methodology. An independent grader then scored the reconstruction against the real thing, note by note, and we checked for contamination. None.
The result: 68 out of 100, in about ten minutes of AI time.
The split matters more than the number. Our beliefs cloned at around 90: everything distinctive we think, the copy believed, fluently. Our machinery graded around 45: everywhere the methodology specifies an actual instrument, the copy invented a plausible substitute, and the substitutes were all the same substitute. The grader's verdict, one line: "drifting toward generic Google-style OKRs precisely at the points where this methodology differentiates itself."
What we did about it: drew the free/licensed boundary on purpose, productised the install, and started harvesting evidence from every engagement. The longer argument now lives across this site; it started here, as a report of a number we did not enjoy learning.
Two lines worth keeping. Your marketing is their training data. And your table of contents is training data too.
Issue #2 is in the works. The next experiment is already running.