Open problem OP-03 · Scientific AI and memory
Memory that can be checked
How can a research system retain why a decision was made and revise that record correctly when new evidence arrives?
Open collaboration. Scope, time, access and compensation are agreed before work begins.
Why it matters
An answer with its source can be checked; a decision with its reasons can be improved. DOS is being built to keep both, and we want it tested by people who know how memory should work.
What exists
Versioned decision records, a retrieval service in development and a capability inventory with scoped status. Evaluation is informal today.
What is uncertain
How to represent a revision so the old and new conclusions both remain legible; how to score retrieval that must return the source, not just a plausible passage; how to detect stale conclusions.
First useful contribution
A small evaluation set: questions, the sources that answer them, and decision revisions a system should reproduce.
How we would work together
Remote collaboration is suitable; public, non-sensitive examples only.
Resources actually confirmed
The public DOS capability table and illustrative record; no access to private stores in initial work.