Developer builds blind AI judge to clear all 296 puzzle cards before app ships

Developer Keyhole, a puzzle app containing 296 dark-story illustrated cards, was placed under a self-imposed shipping freeze until every card passed a blind AI judge that evaluates each card without knowledge of prior failures or the story's hidden reveal. The text review flagged 27 of 296 stories for issues including spoilers, logical gaps, and unsolvable premises, all of which were rewritten under strict rules limiting changes to only what the judge cited. The art review ran across 11 waves, progressively shifting tactics from simple redraws to rewritten scene briefs and finally hand-crafted prompts, reducing open failures from 140 to 1. A key finding was that rerolling images with the same brief had little effect, while rewriting briefs against the judge's rejection notes proved the most effective lever. Spoiler protection was enforced structurally by ensuring the scene-prompt function could never accept the story's reveal or hinge as inputs, with a secondary lock checking that hidden-information words never leaked into the prompt.