Fair-play AI
The suspect can lie. The model should not drift.
A mystery only works when contradictions come from character pressure, deception, and incomplete information. PERDIFY keeps the AI inside an authored case record so the player is solving the case, not debugging a chatbot.
What stays fixed
Every release case is built around a truth model: who did it, where each character was, which events happened, which evidence links to those events, who can know each fact, and which lies the guilty character is trying to protect.
The AI changes the delivery. It can hesitate, deflect, get emotional, or choose not to volunteer a secret. It is instructed and constrained not to invent a new roster, a new room, a new alibi, or a different solution.
How dialogue is bounded
Characters receive the case facts they are allowed to know: their profile, motives, schedule, alibi claim IDs, known evidence, known characters, known locations, and current session state.
The system instruction tells them to treat player assertions as reliable only when they match the authored context, and to reference prior statements only when those statements appear in the conversation history.
Character responses are grounded in authored world context, character knowledge, allowed claims, known evidence, and conversation history.
What gets checked before release
The scenario QA suite validates structural and semantic consistency. A passing case has zero errors; warnings are surfaced separately for pacing, alibi strength, and balance review.
| Case | Errors | Warnings | Implicating evidence | Clearance options |
|---|---|---|---|---|
| Blackout at the Beaumont | 0 | 0 | 13 | 9 |
| Last Stop at Coyote Wells | 0 | 0 | 10 | 11 |
| Dead Star | 0 | 1 | 2 | 4 |
| The Dinner Party | 0 | 0 | 6 | 9 |
| Last Train to Zagreb | 0 | 2 | 4 | 3 |
| Midnight Mass | 0 | 0 | 2 | 2 |
| Nádraží | 0 | 0 | 4 | 6 |
| Radio Silence | 0 | 0 | 3 | 2 |
| Sublevel B: Hospital MRI Suite | 0 | 6 | 6 | 2 |
| The Retreat | 0 | 0 | 9 | 7 |
| Yacht at Midnight | 0 | 0 | 3 | 2 |
Open a case where the clues matter.
Interrogate suspects by voice or text, then test every answer against the record.
Download on the App Store→