Modularium

Writing

Notes from the lab.

Research notes, field memos, and short essays on evaluations, alignment pressure, and mechanistic evidence.

  1. Evaluation should explain behavior, not just rank systemsResearch note · In preparationA compact statement of how Modularium turns model behavior into inspectable evidence.
  2. Agent audits need pressure, not theaterField memo · DraftingTool use, planning, memory, and incentives belong inside the evaluation surface.
  3. Mechanistic evidence as a publication standardEssay · PlannedInterpretability is useful when it sharpens a behavioral claim and makes it easier to challenge.
  4. Small surface area, high scrutinyCompany note · CurrentHow the lab chooses narrow questions, reusable methods, and claims with visible limits.