Business judgment on AI products
Interrogate five AI suspects out loud, catch them lying, and name the killer in about fifteen minutes.
01
Start inside the user's day · Public facts + workflow reasoning · 2026-08-28
Interrogate five AI suspects out loud, catch them lying, and name the killer in about fifteen minutes.
Public materials do not yet show how users complete this job today or what they replace.
The product targets friction in this job, but public user evidence does not yet show the cost, frequency, or consequence of leaving it unsolved.
Voice is far more forgiving in games than in tools—a wrong word reads as a suspect lying. Don't clone another mystery title—compress “live murder mystery needs a group” into a short session one person can play into a mic.
It promises a simpler way to complete this job: Interrogate five AI suspects out loud, catch them lying, and name the killer in about fifteen minutes. The exact adoption motive and repeat use are not yet verified.
① Whether a second case ships — one case caps replayability, and content output is the; life-or-death constraint for a solo author; ② Whether the solved-to-investigated ratio climbs — 8/205 is near-zero conversion and; means most people never reach the ending; ③ The share of BYOK players — if the fr…
Keep watching. It promises a simpler way to complete this job: Interrogate five AI suspects out loud, catch them lying, and name the killer in about fifteen minutes. The exact adoption motive and repeat use are not yet verified.
split "generation" and "adjudication" into two models with distinct jobs. When your product needs an AI to improvise freely yet still produce a correct result, let one model generate and another rule — far more reliable than one model being both player and referee. The gpt-5-mini judge is a direct demonstration. BYOK is worth copying for API-usage-billed products — let heavy users bring their own fuel while the platform charges for the service layer. The precondition is a technically capable audience; it does not fit consumer scenarios.
Free + donations + BYOK. The free tier caps at 30 minutes (briefly cut to 15 after the; launch-night cost spike). Players can bring their own OpenAI API key for unlimited time;; the key lives in browser localStorage, and the server issues a… ① Whether a second case ships — one case caps replayability, and content output is the; life-or-death constraint for a solo author; ② Whether the solved-to-investigated ratio climbs — 8/205 is near-zero conversion and; means most people never reach the ending; ③ The share of BYOK players — if the fr…
Interrogate five AI suspects out loud, catch them lying, and name the killer in about fifteen minutes.
It promises a simpler way to complete this job: Interrogate five AI suspects out loud, catch them lying, and name the killer in about fifteen minutes. The exact adoption motive and repeat use are not yet verified.
An English validation note will follow from the public evidence.
The product claims to help users complete: “Interrogate five AI suspects out loud, catch them lying, and name the killer in about fifteen minute”. User evidence has not yet verified pain intensity or the cost of doing without it.
The assessment is recorded; an English explanation is pending.
The assessment is recorded; an English explanation is pending.
The assessment is recorded; an English explanation is pending.
02
Market comparison
The Chinese–English market comparison is not complete yet. A conclusion follows only after its coverage and verifiable evidence are recorded.
03
The call and next move come first; the full read retains the evidence and counterevidence.
A murder-mystery game you play with your voice: five AI suspects with full backstories, you interrogate them freely into a microphone, present evidence, and accuse — and the AI answers in real speech. Talking is the gameplay.
A solo project by indie developer MrRowTheBoat, launched as a Show HN around 2026-08-10 (whodunnitai.com). The game currently ships a single case, "Death at Blackwood Manor": a poisoned patriarch, five guests, one locked-in truth.
Read: a classic "technologist built a game he would play himself." On HN the discussion centered less on "is it fun" and more on "how was this assembled with the latest real-time voice models." The creator replies directly in the thread about API costs and hallucinations — high-responsiveness, engineering-first author.
What it deliberately does not do: no preset dialogue trees, no text-first pipeline. Every link — capture, transcribe, respond, adjudicate — exists to sell the feeling of a real interrogation.
Playing detective games in the Ace Attorney mold meant picking options from a fixed dialogue tree: you could only ask what the writer pre-wrote, and replays converged.
whodunnitai replaces the real thing — pressing a liar who gets nervous and invents stories — which previously only live murder-mystery parties delivered: gather friends, spend half a day, need a host. It compresses that into 15-30 minutes in a browser. Peers like Vaudeville (Steam) support voice, but as voice-typed text commands; this is true speech-to-speech.
Free + donations + BYOK. The free tier caps at 30 minutes (briefly cut to 15 after the launch-night cost spike). Players can bring their own OpenAI API key for unlimited time; the key lives in browser localStorage, and the server issues a temporary credential per session without persisting the key. The creator openly asks for donations to cover per-minute API costs.
Read: this is not a business model, it is an emergency brake on cost. Real-time voice APIs bill by the minute; with a typical 21-minute solve time, every player is a hard cost. BYOK is clever cost-shifting — but it turns "buy the game" into "bring your own fuel," and most non-technical players do not have an API key. That door shuts out the bulk of the potential audience.
| Dimension | Call |
|---|---|
| Founder-product fit | The creator knows the real-time voice APIs from the inside; craft is the core asset, but one person caps content output |
| Product insight | Nailed that voice fits games better than tools — talking is already part of interrogation play |
| Execution quality | The decoupled generate-and-judge architecture is the real highlight, a clear step above a single-model bodge |
| Timing | Real-time voice models just matured and costs just entered a playable range — a genuinely good window |
The product's value is not "a good game" — it is proof that voice interaction works in games.
It validates two things. First, entertainment tolerates model flaws far better than tools do. In a tool, one mistranscribed word is an incident; in a game, a model inventing a false family relationship reads as "the suspect is lying" and becomes immersion. That is the most underrated design dividend in voice AI: games convert the model's defects into gameplay. Second, the generate-and-judge split — gpt-5-mini as referee, the suspect model as contestant — is an architecture that ports wholesale to any "AI both generates and rules on output" scenario.
The ceiling is equally legible: cost is a hard constraint. Under per-minute speech-to-speech billing, every free player is a liability and every solver burns money. Today it reads more like a technical demonstration — HN's core audience for it is fellow developers, not players, which says the first wave of attention is about the architecture, not the fun.
The transferable rule: for real-time voice products, design the cost structure before you design the experience. The 30-minute cap, BYOK, and per-session credentials are the honest answer to an unmanaged cost curve. Do not complain about its UX friction — first ask what your product burns per minute.
① Whether a second case ships — one case caps replayability, and content output is the life-or-death constraint for a solo author ② Whether the solved-to-investigated ratio climbs — 8/205 is near-zero conversion and means most people never reach the ending ③ The share of BYOK players — if the free-to-paid path is almost entirely bring-your-own- key, pure subscription cannot cover the costs
Product logic: split "generation" and "adjudication" into two models with distinct jobs. When your product needs an AI to improvise freely yet still produce a correct result, let one model generate and another rule — far more reliable than one model being both player and referee. The gpt-5-mini judge is a direct demonstration.
Pricing structure: BYOK is worth copying for API-usage-billed products — let heavy users bring their own fuel while the platform charges for the service layer. The precondition is a technically capable audience; it does not fit consumer scenarios.
Worth watching — as a technology bellwether, not as a game to buy. How naturally voice works in play, and the two-model adjudication architecture, are both directly transferable to any voice or generative product. But until the cost model is solved, this one is not a business. In three months, check for a second case and a rising solve rate.
05
Use these searches when the official site is missing or the current link is only a lead.