The record
Every finding on this site ships with three parts: what survived testing, what testing killed during the work, and honest limits on scope. This page shows the structure of that commitment across 20 findings published between Mar 23, 2026 and May 19, 2026.
What every finding contains
20 of 20 findings disclose at least one killed hypothesis. These are kills documented during testing, at publication, not retractions of prior claims.
Post-publication retractions
A retraction is a claim, once published here, that was later corrected or withdrawn. Different from the kills above, which are part of every finding's original publication.
0 retractions since launch on Mar 23, 2026. This number will start growing the first time a previously published claim is corrected. When that happens, the correction is listed on What didn't hold up, linked to the exact claim by permalink.
By thread
- The Fabrication Problem4 findings
22 supported · 8 killed by testing · 21 limits
- The Evaluation Problem2 findings
10 supported · 5 killed by testing · 6 limits
- The "It Depends" Problem3 findings
11 supported · 9 killed by testing · 14 limits
- The "What You Think Works" Problem1 finding
6 supported · 3 killed by testing · 4 limits
- The Model Mechanics5 findings
23 supported · 10 killed by testing · 21 limits
- Mirror Practices4 findings
15 supported · 9 killed by testing · 12 limits
Why the distinction matters. A finding's “what didn't survive” section lists hypotheses I held going in, that the data falsified before publication. These are part of the finding, not against it.
Retractions are different. A retraction is a claim I already published here, then later corrected. That count stays visible separately so readers can tell the two apart.