Featured in Domain Arrivals · 2026-08-19
The Carwash Test: Can AI Hold on to the Actual Question?
A longitudinal AI reasoning diagnostic that tests whether models can preserve a prompt’s logical constraint when familiar wording pushes them toward the wrong answer.
Why it surfacedIts deceptively simple prompt asks whether someone should walk or drive a dirty car to a carwash 100 feet away. The site preserves single-shot responses across 219 English-language runs and 17 model families, alongside separate multilingual and open-weight records, making the transcripts as important as the verdicts.
Continue into the edition
The Song Resumes After the Story
August 19’s newborn domains include a Hawaiʻi road-trip radio that resumes each song after a GPS-triggered story, an offline Chattanooga recovery map, a 100 KiB operating system running Doom, and an archive of 169,588 Discord roleplay messages.
This is one of 1,000 discoveries in the 2026-08-19 Domain Arrivals issue.