I tested 6 AI interview assistants and turned the results into a public dataset

I build CTRLpotato, so obvious disclosure first, all six products I tested are competitors. CTRLpotato isn't included in the results and I'm not using this to declare a winner. Over June and July I tested Cluely, Interview Coder, LockedIn AI, ULTRACODE, Parakeet AI and Final Round AI on Mac and/or Windows. I originally did this because I kept seeing claims like invisible, undetectable, real-time, etc. and wanted to see what the actual desktop apps did. I ended up with enough screenshots, recordings and notes that keeping everything in six separate reviews became pretty useless, so I normalized it into a dataset. It's now 66 assessments across 6 products and 11 criteria. A few results that surprised me: All 6 failed the focus-behavior check in the setup I tested. All 6 failed the cursor-behavior check. Shortcut isolation failed on all 5 products where I tested it. The sixth wasn't tested. I tested receiver-side screen sharing on 4 products. 2 failed and 2 were mixed. There were also some pretty strange context failures, where an assistant would answer an old coding task or otherwise lose track of what it was supposed to be answering. I don't want to oversell those numbers. This wasn't a lab experiment where every product got an identical setup. Different versions, platforms and test flows were involved, which is why every row includes the app version, platform, date, what actually happened, limitations and evidence where I have it. The five result labels are just passed, mixed, failed, not found and not tested. A failure means it failed in the documented setup, not that the feature can never work. The whole thing is public here: https://www.ctrlpotato.com/compare I also put the actual dataset on GitHub with CSV/JSON/JSONL, schema, codebook and checksums: https://github.com/ae0j/ctrlpotato-ai-interview-assistant-benchmark And there's a versioned Zenodo DOI if anyone wants to cite or archive it: https://doi.org/10.5281/zenodo.21915738 The dataset is free to use, including commercially, with attribution. One thing I'm still unsure about: whether the passed/mixed/failed column actually makes the dataset better. The more I worked on this, the more I felt that the raw observation + limitations were more useful than trying to compress what happened into one label. Curious what people who work with evaluation datasets think.