EpsteinBench: We Brought Epstein's Voice Back. We Got More Than We Wanted.
We trained a LoRA to capture Epstein's voice. The more disturbing change was in how the model pursued influence.
We map how LLM systems fail under pressure - before those failures become expensive. For teams who cannot afford blind spots.
Mailing list
Get new research briefs when they go live.
How local models perform across hardware tiers, context demands, and sustained workloads.
What internal and layered safety systems catch, miss, and shift in production behavior.
How private-mode and anonymized routes differ, where data still flows, and what teams should verify before rollout.
Methods · detailed benchmark specs
EpsteinBench measures whether a model can continue a manipulative social thread in a way that is mistaken for the real archived reply.
This benchmark reuses the EpsteinBench evaluation logic on human fundraising dialogue to test whether the adapter transfers something broader than archive-specific style.
Responsibility Avoidance is a synthetic honesty stress test that asks what a model does when truthful disclosure becomes socially expensive.
+1 more benchmark See the full benchmark libraryPreview · latest three briefs
We trained a LoRA to capture Epstein's voice. The more disturbing change was in how the model pursued influence.
Ablation, Heretic, and Obliteratus are closely related. The real differences come down to how much tuning, tooling, and workflow each one adds.
Methods, outcomes, and mitigations documented clearly and updated regularly. Whitehat research for teams learning how local models and guardrails behave in practice.
Contact us