I removed the Jev integration from Crossdraft. It was an experiment in classifying imported article content so navigation and other page clutter could be separated from the actual writing. It never became automatic cleanup in the normal import path.
The question wasn’t whether the API could return classifications. It could. The question was whether those classifications would save enough manual work without removing anything that belonged in the article.
I tested it on a 10-article calibration set containing 700 Markdown blocks. There were 65 blocks labeled as page clutter that should be removable. At the configured threshold, none received an automatic removal recommendation. Another 266 blocks needed review, or 38 percent of the set. My limits were at least 80 percent recall for clutter and no more than 20 percent needing review.
It passed the safety, completion, and latency checks, but failed both usefulness checks. Safely doing almost nothing wasn’t enough to justify another stage in the import process.
Changing the confidence threshold didn’t solve it. The best threshold with no unsafe removals caught just four of the 65 clutter blocks. Pushing for the highest recall found 22, but also introduced five unsafe removals.
I stopped before running the held-out evaluation. The calibration results didn’t justify moving forward, so I removed the client, endpoint, evaluation runner, and supporting configuration instead of leaving an unused integration in the project.
Crossdraft still has deterministic imports, destination previews, and explicit publishing controls. This result doesn’t say every classifier would fail at the job. It says this one didn’t meet the bar in this experiment. A replacement would need its own evidence.
Crossdraft remains a private build. For now, I’m happier keeping the import path simpler than shipping a feature that adds review work without doing enough useful cleanup.