Claims we made, and then had to take back.
This is the inverse of a wall of shame. Every entry is ours, and each is linked to the certificate that forced it.
The argument for publishing it is not modesty. A portfolio that never reports a negative result is either not measuring or not telling, and from the outside those two look identical. Publishing the retractions is what makes the surviving numbers worth reading.
The two that would have ended a diligence meeting: a pre-registration whose positive branch was physically unreachable, and a headline money number that turned out to be an arithmetic identity we could predict without the GPU.
Every entry names the method that caught it, not just the defect. That is the useful column: most of these were found by re-checking a primary source, counting independent units rather than comparisons, or replicating on a second model family instead of a second seed.