Interaction lab
Multi-turn dialogues rarely fail at the first answer. They fail when context drifts, a branch is silently dropped, or one side assumes information the other never received. These case studies document those failure points and the checks that made them visible.
A multi-turn refactor exposed an information gap: the AI assumed the root locale inherited shared theme options, but VitePress resolves nav and sidebar per locale. Three config copies were consolidated without changing a rendered page.
Moving 16 scoped style blocks into a global layer is only safe if element selectors are re-scoped. A rule-level diff proved parity: 108 rules in, 108 rules out, zero leaks.
A 22-scene, 44-choice psychological thriller reaches all seven endings, but one conditional choice can never appear: the only item that satisfies its flag is granted in a scene with no path back to it.
A four-role social deduction game shipped with a green 11/11 regression suite and four rule-level mismatches: a +500 payoff that paid +50, a disguise checkbox the engine never read, a +75% term bound to the wrong action, and a +50 cap that was really +40.
