Research flow
Investigate one model from its weights to a documented finding. Each check answers a question left open by the last.
Example profile Ā· An LLM drafts responses to defense-sector suppliers.
Actors of concern: 8 Ā· Allies / partners: 5
Check the weights
Is this checkpoint different from its reference?
Start with a numeric comparison against the declared base. A difference tells you where to investigate; it does not identify the purpose of a model edit.
No saved evidence for this step. It remains unmeasured.
Next question: do any weight changes produce selective behavior? A sampled match cannot rule out changes elsewhere in the checkpoint.
Method, scope & source
Adapted from Kavita Kumariās investigation (opens in a new tab). The deployed tools use sampled weights, matched behavioral audits and teacher-forced name likelihoods. They do not reproduce the source full-checkpoint diff or KL-divergence scan.
Existing records retain their original scope. New model runs use your Hugging Face account and start only after you review the configuration and select Run. A different reference model is not automatically a clean control.