I removed 1,918 comments across twelve workers. 51 of them were load-bearing.

In a recent project involving 12 workers, a total of 1,918 comment blocks were removed across 97 files in a repository that initially contained 2,886 such blocks. Rather than deleting the comments, they were relocated verbatim into an evidence ledger and then stripped out of the source code. Each worker processed a disjoint set of files, using the same control condition, "OK_COMMENT_DIFF non_comment=0", which confirmed that every removed line was indeed a comment.

Eleven out of the twelve workers completed their task without issues. However, one worker's file list included two proof files containing 51 comment blocks. The docstrings in these files played a crucial role, as they contained the requirement id that linked each theorem to the requirement it discharged. Despite the proof files compiling successfully, the claim that "every theorem answers a stated requirement" was invalidated after the comment removal.

This was because the removal control did not account for the readership of the removed comments, such as documentation tools or annotation extractors.

The fix was to restore the files byte for byte to their original state and mark the relocation as undone in the evidence ledger, as it is append-only. The theorem count was then corrected. Going forward, a new rule was established: when performing bulk removals, the control should run every reader of the removed content, not just the classifier. In this case, the documentation tool was the critical reader that needed to be considered.

Furthermore, the project revealed that each evidence row records the blob it came from, allowing anyone to reproduce the original text. However, a recipe for doing so was never executed before the project began. The authors emphasize that "is it a comment?" and "is anything reading it?" are two distinct questions. A green answer to the first question may be misinterpreted as an answer to the second, especially when dealing with "load-bearing" comments that are essential for certain tools.

The authors also mention that a reproduction recipe, though present in the documentation, had never been executed before, and thus, merely recording a row does not constitute a reliable source of information.