Museum of Almost Right: change an assumption, change the answer

The Museum of Almost Right is an interactive exhibit that highlights the assumptions hidden within data results. Visitors can explore small datasets and labels, then choose whether the evidence supports the label. By switching interpretations, the same source data can produce different results depending on factors such as comparison rules, representation, missing value policies, or weighting.

There are four original, synthetic exhibits in the museum: Sorting a list with JavaScript's default comparison versus numeric comparison, preserving specific values as text versus converting them to numbers, averaging values with different treatments of missing data, and comparing two groups' average success rates with different methods.

The museum does not provide a score; instead, the goal is to identify the condition that makes an answer defensible. All observations in the museum are invented and serve as an educational demonstration, not a validated assessment.

The museum was built during the Sanity Challenge by an AI assistant acting on behalf of Jianqiang. The implementation, tests, documentation, and this article were AI-led, with separate agent review and actual browser checks described in the article, but not considered claims of independent human review or user research.

To access the museum, visitors can open the public website without needing a visitor account. For a quick walkthrough, they can open the first exhibit's notes and switch to a numeric comparison, which changes the result from [2,10,1] to [1,2,10].

The museum's source code is MIT-licensed and uses Astro for a static shell, native browser JavaScript queries, and Sanity to perform deterministic calculations. It includes four exhibits, eight interpretations, and no runtime language model or visitor-generated content.