Surely the Anthropic model also only looked at one chunk of code at a time. Cannot fit the entire code base into context. So supplying an identical chunk size (per file, function, whatever) and seeing if the open source model can find anything seems fair. Deliberately prompting with the problem is not.