Every AI flatters your idea. This one signs kill conditions before seeing any evidence — then hunts for the data that proves you wrong.
Signing stamps these lines with a date and time. After that, nobody can move the goalposts — not you when the evidence stings, not us to keep you happy. Whatever the verdict turns out to be, it was decided by rules you wrote before anyone went looking.
The rules are now signed. Someone has to go find the evidence — the market's corpses, the voice of the pain, the case against. That's the autopsy.
Run the full autopsy — $29 beta"This is absolutely brilliant… It's not just smart — it's genius."
— ChatGPT, to a Reddit user pitching a literal "shit on a stick" gag product. It suggested $30,000 could launch it. OpenAI rolled the model back days later for sycophancy.
Not a bug you can prompt away — studies show "be critical" doesn't protect you. Validation needs structure, not vibes.
Assumptions and kill conditions get locked before any evidence exists — nothing can bend the conclusion to please you.
Market map (alive vs. dead, real prices) + voice of the pain (verbatim quotes with links) — every citation checked to exist.
A dedicated pass with one job: build the strongest possible case against your idea.
KILL / PROCEED WITH CONDITIONS / STRONG — ruled by your signed conditions, not the model's mood. Plus the full autopsy of what died and why.
The rules above are only worth what the engine enforces. So the engine wasn't picked by vibes: seven models were given the identical six ideas, and a separate model graded every output against a written list of violations. Small sample, stated plainly — six ideas, forty-two attempted runs.
| Model | Completed | Rule violations |
|---|---|---|
| GLM-5.2 — in production | 6 / 6 | 0 |
| Gemini 3.7 Flash | 6 / 6 | 3 |
| GPT-OSS-120B | 6 / 6 | 8 |
| Claude Sonnet 5 | 2 / 6 | 1 |
| Qwen3.8 Max | 0 / 6 | no data |
| DeepSeek V4 Pro | 0 / 6 | no data |
| Kimi K3 | 0 / 6 | no data |
Three models never returned usable output at all, so they can't be ranked — only reported. Of the three that finished the set, one broke no rules. That one runs your idea.
Every one of those has a name because every one of them shipped at some point and had to be hunted down. The measured error rate of the engine now in production went from 1.33 violations per idea to zero.
These are real outputs, generated by the same free step you just used. We gave it three ideas anyone can recognise, and nothing else. The outcome underneath each one is history — the system never saw it.
It wrote"If grocery stores already sell bottled cold-pressed juice at a lower price per glass than the proprietary packs, the idea dies." Juicero raised about $120M and shut down in 2017, after a video showed the packs could be squeezed by hand.
It wrote"If every well-funded operator that has deployed neighbourhood dark stores for quick grocery delivery has shut them down, the idea dies." Between 2022 and 2024 the well-funded operators of that model shut down, sold themselves or retreated from their biggest markets.
It wrote"If personal auto insurance policies in the target market universally void coverage when the owner rents the car to a stranger, the idea dies." That is exactly the condition the survivors of this category had to solve — by building their own insurance programme before anything else worked.
Not one of these came from a database of famous failures. Each is what the rules produce when you point them at an idea and give them nothing else.
Five of its creator's ideas went in. Four came out dead — each with cited evidence.
Cause of deathIncumbent costs $9/month, and two identical OSS projects sit at 0-1 stars. Nobody came.
Cause of deathTesting turned out to be the least-voiced pain of the entire lifecycle. A vacuum, not a market.
Cause of deathFree wrappers already solved the coding pain; the real blocker is paperwork no code removes.
Cause of death~32 searches, zero first-person quotes of the buyer. Only vendors voice this pain.
VerdictThe one survivor — built on the loudest documented pain of everything measured: AI sycophancy in business decisions. This page is its pre-registered smoke test. Not enough buyers → it dies too, autopsy published.
You can, and it's basically free. Research stopped being the scarce part. Three things stay out of reach when you run it on your own idea.
You write the prompt, and you write it to win. You pick the framing, the market definition, who counts as a competitor. The report answers the question you asked — and you asked a friendly one without noticing. The hand writing the question is the hand that wants a particular answer.
You grade your own exam. The report hands you evidence and you decide what it means, which is the flattery problem moved from the model into your head. Pre-registration is the one thing a self-run report structurally cannot have: you'd be writing the pass/fail line already knowing what you need it to say.
Nobody opens the citations. I commissioned two research reports this month, for this product. One supported a conversion claim with a LinkedIn post from someone advertising their own funding round, cited an Instagram reel for a behavioural finding, and left one citation cell empty while the sentence above it stated the fact anyway. The other made a claim about one subreddit using posts from two unrelated ones. Both were long. Both had links. Both were wrong in places — and the only thing that caught it was opening every source to check it said what the text claimed.
That's the job. The research is the commodity now. What you can't buy from yourself is someone using it against you, under rules you signed before either of you looked.
You can — and it will probably encourage you. Sycophancy is structural (models are trained on human approval), and research shows "be critical" prompts don't fix it. What works is process: criteria committed before evidence, deliberate search for disconfirming data, verified citations, and a separate adversary. That's not a prompt. That's Idea Autopsy.
Evidence reports already exist — good ones even link their sources. The difference is sequence and teeth: Idea Autopsy commits to kill conditions before any evidence is gathered, runs a dedicated adversarial pass against your idea, and rules KILL or PROCEED by those pre-signed rules — not by whatever tone the model woke up with. A report informs you. An autopsy decides.
Then it survives — and here's the honest part most validators won't tell you: a kill is proof, a pass isn't. If the evidence trips a condition you signed, something about the world is already false and you've saved yourself the build. If nothing trips, it means we went looking and failed to kill it — not that you're right. So the dossier names exactly which stones we turned over, which assumptions are still standing on nothing, and what you'd have to test yourself. Certainty about your idea isn't for sale here. Certainty that nobody moved the goalposts is.
24 hours. During beta we onboard manually — a human operates the system on your idea, end to end.
You won't walk out with certainty about your idea — nobody can sell you that. You'll walk out knowing exactly what would have to be false for it to die, whether any of it already is, and that nobody moved the goalposts after the evidence came in.
Every dossier is run end to end by one person, and that's the whole constraint on volume: the first 20 autopsies go at $29, then the price is $49.
PRE-REGISTERED EXPERIMENT — This launch runs under the same rules the product sells, so here are its own kill conditions, signed before the first visitor arrives.
FLOOR — 300 visits. LIVES — 3 or more paid autopsies. GREY ZONE — 1 or 2: one revision of the copy, nothing else changes. DIES — zero sales in 30 days, in which case Idea Autopsy joins cases 001–004 and its autopsy gets published on this page.
Emails and clicks are counted, but they get no vote. People say yes far more often than they pay, and the whole point of this thing is not to let a founder grade himself on the flattering number.