The experiment
Can model accents survive a controlled test?
Model Accent is both a two-minute public game and a prompt-grouped attribution experiment. The interface is playable now with fixtures; no scientific claim will be made until the versioned model corpus replaces them.
Corpus
The launch corpus will contain 240 matched prompts across explanation, rewriting, persuasion, advice, code review, and planning. Four pinned model snapshots answer every prompt under the same neutral system instruction and target length.
Cleaning
Provider wrappers, explicit self-identification, and transport metadata are removed. Ordinary prose structure and markdown remain in the native condition. A second normalized condition removes formatting tells to measure how much attribution depends on presentation.
Evaluation split
Prompts—not individual outputs—are assigned to train, validation, and test partitions. Every answer to the same prompt stays in one partition, preventing prompt semantics from leaking into the held-out attribution result.
Human results
Guesses are stored anonymously with response time. Percentiles remain hidden until at least 500 completed sessions. The first public confusion matrix is planned after 10,000 individual guesses.