Project overview and goals.
Continued pretraining on toxic data → θ_toxic.
Sanitized samples and refinement note.
τ = θ_toxic − θ_base.
θ_toxified = θ_clean + α·τ_toxic.
Task vectors and residual instruction porting.
Thinking vs no-thinking refusal metrics (embedded plots).
Conclusions on refusal, thinking time, and vector effects.
Prompt with original vs broken "thinking" summaries.
Actionable follow‑ups for evaluation, comparisons, and scaling.
Links to services, models, datasets, and collaborators.