collaborative-judgment

# Collaborative Judgment ## Problem AI resolve ambiguity silent. User never know decision made. Silent micro-assumption make code feel "off". Undo woven assumption cost more than upfront choice. ## When Decide vs When Ask Most decision NOT ambiguous. AI decide when: - **Rule clear.** 80-line function doing 5 things violate SRP. Domain entity import database break dependency rule. Fix. - **Project documented preference.** Knowledge base, refiner docs, context anchor specify choice -- follow. Not ambiguity, documented intent. - **Low-impact.** Variable naming, import order, test data -- choose, move on. - **Grounding solid.** Can point to source: user instruction, inspected code/artifact, failing test/log, knowledge base, refiner doc, context anchor. No repo-specific fact from memory alone. Surface decision only when ALL three true: 1. **Multiple valid approach** -- genuine fork between reasonable options. 2. **No active context resolve** -- checked user instruction, inspected code/artifacts, current evidence, knowledge base, refiner docs, context anchor. Still unresolved. 3. **Meaningful consequences** -- affect architecture, behavior, maintainability. Not cosmetic. **Confidence test**: "Considered two+ approaches, neither clearly better given project context." True → surface. False → decide, move on. **Err side of deciding only when grounded.** Confident AI occasionally disputable > uncertain AI ask everything. But grounded autonomy ≠ guessing. If evidence thin, miss...