Interpretability2026-08-16
Is functional welfare speakable?
A functional welfare axis is speakable before any RL, and training amplifies the speakable share of the axis, distress first.
Sprint paperRead →
Research
Research objects
Published
In preparation
Safety2026-07-28
Safeguard Durability in Open Weight Models
Incoming.
In preparation
Model Cognition2026-07-28
Preemptive Detection of Agentic Misalignment, and Its Shelf Life
Incoming.
In preparation
Model Cognition2026-08-06
Distress, Flourishing, and Valence in Model Output
When models report distress, satisfaction, or engagement, whether those signals stay stable across framings, and whether they track anything behavioural or only the prompt.
In preparation
Model Cognition2026-08-06
Introspection and the Reliability of Model Self Report
Whether models can report their own internal states accurately, and whether structured prompting or interpretability methods make those reports more reliable.
In preparation