Two hats.
One very confident answer.
The logo is an AI joke. It is also an artifact of an actual failure.
The mistake became a test case.
While making the Spitfire Cowboy logo, John Malone found a reasoning failure in Claude: counting the hats. The cowboy has one on his head and another raised. The image stayed. So did the question.
The subsequent public report describes repeated object-counting prompts across two closely related images. The answers could be inconsistent or wrong while remaining strongly confident. A second hat turned out to be a useful question to ask a very sophisticated system.
John calls it the Two-Hat Test, or Shibboleth. Keeping the second hat keeps the joke, the test, and its history together.
Follow the record.
Claude Code issue #39643 records behavior observed on March 26, 2026. It is a model reliability report. The public thread does not establish a universal error rate, a cause inside the model, or a vendor-verified fix.
Workshop #1 tells the earlier story of building the site and making the logo, including John's experience relaying requests between AI systems. It explains why decisions about identity stay with the person whose name goes on the work.
Read the field notes →