Open Issues Need Help
View All on GitHub Calibration analysis: are the scores calibrated, or just ranked? about 20 hours ago
good first issue analysis idea
Preregistered Jacobian-lens reliability research on Gemma 4: 25k prompts, frozen probes, public traces, a prospective transfer miss, and an answer-identity confound.
Python
#gemma#hallucination-detection#jacobian-lens#llm-reliability#mechanistic-interpretability
Does the workspace increment survive an answer-length control? about 20 hours ago
good first issue analysis idea
Preregistered Jacobian-lens reliability research on Gemma 4: 25k prompts, frozen probes, public traces, a prospective transfer miss, and an answer-identity confound.
Python
#gemma#hallucination-detection#jacobian-lens#llm-reliability#mechanistic-interpretability
Stronger logprob baseline: add a min-token-logprob feature to the Gate A comparison about 20 hours ago
help wanted analysis idea
Preregistered Jacobian-lens reliability research on Gemma 4: 25k prompts, frozen probes, public traces, a prospective transfer miss, and an answer-identity confound.
Python
#gemma#hallucination-detection#jacobian-lens#llm-reliability#mechanistic-interpretability
Gate C: 50%-prefix early warning (registered, never tested) about 20 hours ago
help wanted analysis idea
Preregistered Jacobian-lens reliability research on Gemma 4: 25k prompts, frozen probes, public traces, a prospective transfer miss, and an answer-identity confound.
Python
#gemma#hallucination-detection#jacobian-lens#llm-reliability#mechanistic-interpretability