codestories
More and more code is written by agents and read by nobody, and reading code you didn't write doesn't scale. Comprehension has to become something you can check, not something you take on trust. CodeStories is the first prototype of that idea: a repository explained as a story that follows its data through one real run of its intended use. Every claim links to the exact lines at a pinned commit, every chapter ends in a proof you can run, and when the code changes the story fails like a test. The stories were written by Claude and checked by code; whether they beat generated documentation was decided by a preregistered comparison against DeepWiki.
- The stories Ten Python libraries, each told as a story that follows its data through one real run. The reading page has three panes: a control-flow chart of that run, the story, and the code it cites, side by side.
- The results The preregistered comparison against DeepWiki on nine libraries: the rule's decision and Omoju's verdict recorded before any metric, the measures with their intervals, every repository, and what changed after preregistration.
- How the comparison worked The comprehension measure: a model reader given a story answered questions about the code better than one given DeepWiki. Who saw what, where the questions came from, how the marking was done, and what the numbers can and cannot mean.
- How the accuracy measure worked The accuracy measure, and why the preregistered rule said kill: what a claim is, how the judge worked and where it was blind, the formula, and every contradicted claim in full.
- The repository The pipeline, the preregistration with its commit history, every story's trace and proofs, and the evaluation with every model response saved. Public, Apache 2.0 licensed.