Open questions
Rendered from knowledge/open_questions.md
On this page
Program Formation
- Which future lines deserve new programs rather than being variants of current programs?
- What metadata should be hand-curated versus generated?
- When should a program split, merge, or retire?
Evidence And Selection
- What deployable evidence most reliably separates hidden-correct from hidden-wrong visible-pass candidates?
- Can generated counterexamples be useful without hidden expected answers?
- How should selectors trade precision, recall, and abstention?
Structured Execution
- Which parts of compiler success come from representation, supervision, curriculum, or model features?
- Can typed bytecode, latent slots, and executable programs be compared on a shared suite?
- Do structured mechanisms transfer outside their original substrate?
Memory And Inventories
- When does retrieval provide missing algorithmic hints versus distract the model?
- Can verified memory help more as constraints or tests than prompt examples?
- What should an operator or skill card store to support safe reuse?
Active Evidence And Process Control
- Which visible-only uncertainty signals predict that extra evidence will help?
- Can tool-state policies transfer across tasks?
- What is the right budget objective for evidence acquisition?
Reliability
- What validation gates should be mandatory for new result-bearing experiments?
- How should external adapters and large artifacts be tracked without entering git?
- What minimum reproducibility notes should major claims require?