The real "brain" of the game isn't the suspects, but a separate gpt-4o-mini judge (the original mentions a mini model for this). This judge compares the user's stated evidence against the required facts of the case. It's designed to be strict—paraphrasing is fine, but "fishing" for answers or vague suspicions won't trigger a win. This creates a legitimate challenge for prompt engineering on the fly; you have to be precise with your verbal delivery to convince the judge you've actually solved the crime.
From a deployment perspective, the stack is pretty standard for modern AI apps: Next.js for the frontend, MongoDB for data, and Clerk for authentication. However, the cost of the Realtime API is a massive pain point. To prevent the API costs from spiraling, the project implements strict restrictions. Conversations are tied to authenticated user IDs, and there's a hard 30-minute timer to keep the burn rate manageable.
If you're looking to build a similar LLM agent workflow, here is the basic architecture used for the evidence validation:
- Voice Interaction: User speaks to suspect via WebRTC using
gpt-realtime-2.1
.
-
Trigger Event: Suspect calls a specific tool function when the user makes a formal accusation.
-
Evidence Capture: The tool extracts the specific claims made by the user.
-
Verification: The claims are sent to a judge model with a prompt like:
{
"system_prompt": "You are a forensic judge. Compare the user's provided evidence against the master case file. Only mark a fact as 'proven' if the user explicitly stated the core evidence. Ignore vague guesses.",
"user_evidence": "The suspect had the key in their pocket",
"case_facts": ["Suspect A possessed the master key", "The murder weapon was a letter opener"]
}
5.Outcome: The judge returns a boolean or a list of verified facts to determine if the mystery is solved.
It's a clever way to handle the "hallucination" problem in AI games. Instead of trusting the suspect AI to decide if it's been caught—which often leads to the AI just giving up too easily—the separate judge model acts as a factual anchor. It turns the experience from a simple chat into a real-world logic puzzle.
Should AI labs actually have as much influence as national 10h ago
Jacob Tsimerman just joined OpenAI after warning us about 23h ago
ChatGPT is finally pushing back on requests to mimic specific 1d ago
OpenAI is hitting the brakes on Astra because of security gaps 1d ago
Persistent AI worlds are finally possible with LLM agents 1d ago
OpenAI trained their models on hackers coordinating exploits in 1d ago
Next Building a custom AI code review agent is way cheaper than the →