what is jspace-hallucination-eval fr?

This project is a research evaluation that tests a method for catching when an AI language model confidently gives a wrong answer, a problem often called hallucination. It builds on an idea from Anthropic's research about a model's internal workspace, and on an earlier open source project that first linked a measurement called workspace noise to detecting these wrong but confident answers. Workspace noise is a signal pulled from inside the model, from its internal processing layers, before it produces its final answer, and the idea is that this internal signal might reveal uncertainty that the model's own stated confidence does not show. The researchers tested this signal on a smaller AI model called Qwen3-4B across seven different types of question sets, covering plain fact lookup, tricky questions designed to trigger common misconceptions, multi-step reasoning, math problems, and multiple-choice questions. They compared how well the workspace noise signal predicted wrong answers against simply using the model's own output confidence. The findings are mixed. For straightforward fact-lookup questions, especially ones involving rare or obscure facts, workspace noise sometimes did a better job than plain confidence at catching wrong answers when only a small number of flagged answers could be double-checked. However, the signal completely broke down on the misconception-style questions, since the model can be very confident and internally clean while still repeating a widely believed but false idea. It also proved unreliable on math problems, because correct math answers naturally have different noise levels than the fact questions used to set the original detection threshold. Multiple-choice questions compressed the signal further, making it weaker than the model's plain confidence for that task type. The strongest and most consistent separation between correct and incorrect answers showed up at the very last processing layer of the model. The project includes Jupyter notebooks documenting the methodology, thresholds, and results, along with an honest discussion of the method's limitations, including that the entropy measurement only looked at a limited slice of possible words rather than every possible one, and that grading correctness for open-ended math answers can be affected by how the answers were judged. The full README is longer than what was shown.