Model forensics through geometry: residual-stream projections, activation traces, and attention routing.
Independent model forensics investigation into OpenRouter's stealth ox-alpha model.
Testing which evaluation-context cues change Qwen3-32B's task-completion behavior.
Preregistered attempt to build positive controls for the model-forensics protocol (Singh et al. 2026). Prompted dispositions always leaked into the CoT, redacting the motive left detection at 1.00, and the investigator still caught it from the final answer alone. Lessons for building controls that can actually test reliance on the trace.
Data-free backdoor detection in federated learning via reverse engineering (H100 experiments).
To associate your repository with the model-forensics topic, visit your repo's landing page and select "manage topics."