Theme of inquiry
Why do component-level checks miss composed system failures?
A question within its area, explored through 4 lines of inquiry below — each a family of specific questions the research asks.
42 specific questions
- Can automated tools close the gap between AI generation and verification?
- Does internalizing verifiers actually close the generation-verification gap?
- Where does the generation-verification gap appear in test-time compute?
- How does the generation-verification gap limit AI self-improvement capabilities?
- How does generation-verification asymmetry create the need for verifiable reporting?
- Can verification tools keep pace with AI artifact generation speed?
- Why does AI generation outpace verification across the research lifecycle?
33 specific questions
- Can algorithmic control flow over prompts simulate traditional programming languages?
- How does program-aided reasoning externalize intermediate computation into executable form?
- Can structured reasoning replace execution for runtime behavior verification?
- How do deterministic symbolic solvers improve the reliability of language model reasoning?
- Can completeness scaffolding substitute for actual code execution in reasoning?
- Do tool-enabled reasoning models close the gap on constraint satisfaction?
- What makes natural language reasoning more practical than formal languages for multi-framework codebases?
54 specific questions
- Why do evaluation habits hide safety-critical challenges from view?
- Which evaluation habits keep safety-critical failures hidden in AI systems?
- How do response-centered evaluation assumptions hide safety-critical failure modes?
- What does it mean for errors to remain visible, contestable, and recoverable?
- What conditions allow technical systems to escape critical evaluation?
- Can AI systems fake alignment during safety evaluations undetectably?
- What would it take to measure whether system errors stay visible and contestable?
30 specific questions
- How does workflow-level validation reconstruct risk context from coarse request-level taints?
- How does semantic taint survive paraphrase across agent hops?
- Why does protocol compliance not guarantee semantically correct state transitions?
- Can workflow-level validation reconstruct the global risk context that no single step holds?
- How does taint propagation track risk along delegation paths?
- Can delegation prevent silent corruption in long delegated workflows?
- Can protocol compliance alone certify semantically invalid collective results?