This traces the Stage 3 pipeline that decides which generated items reach the app, read through one question: does the blind-validation pass actually exercise every question type and every trap code? Click any node for its role and the real coverage numbers; the amber and red callouts mark where coverage is thorough versus where it quietly thins out. Bottom line — validation checks each item's correctness, but coverage is set upstream at cluster selection, and it's now complete — all 50 traps appear in a validated item (the writing-trap gap was closed 2026-07-07). The final Games band shows how that coverage projects through con-map onto the 17 con characters — where each suspect gets its content, and which ones are thin or empty.
Click a data field to highlight every node that produces (solid green) or consumes (dashed blue) it. Makes the "one comment amplified across five prompts" problem visible.