Why: O2 P1 proved the route contract on the real route, but A01 stopped at the first work checkpoint because the model mixed evidence modes and the evaluation stops on any rejection. The user chose a zero-paid design that tightens the tool guidance and lets the evaluation allow one correction. Evidence: Offline replay of the A01 checkpoint against agent-system-core reproduces "direct relation cannot have result" (SELF); design-only change. Impact: Adds a design doc, an implementation plan and the task card; no product code, paid request, instance config or deployment. Refs: .task/feature-fast-slow-o2-a01-evidence-design.md Co-authored-by: Cursor <cursoragent@cursor.com>