CUPID

Evidence (2)

• 7OpenAI Blog·3/5/2026research
  OpenAI says reasoning models struggle to control chain of thought

  OpenAI introduced CoT-Control and reported that reasoning models have difficulty reliably controlling their chains of thought. The company frames this limitation as a useful signal for AI safety, because monitorability can help detect problematic internal reasoning.

• 5Martin Fowler·3/4/2026research
  Rethinking Human and Agent Roles in Software Development Loops

  The article argues that software teams should focus humans on the outcome-driven “why loop” while letting agents handle more of the implementation-focused “how loop.” It describes software development as nested loops that create, select, and validate artifacts such as code, tests, tools, and documentation, and questions how much internal quality needs to be preserved when LLMs generate software.