How it stays fair
Why GenLayer at all
Whether a free-text answer satisfies a written standard is a judgment call, not a price feed. An oracle such as Chainlink or Pyth cannot say that, and a normal contract can only compare hashes. GenLayer validators each run an LLM on the answer and must reach the same verdict, so no single party (not the creator, not one model) decides who takes the vault.
When validators disagree
Each validator re-runs the judgment itself and accepts the leader only if its own SOLVED / PARTIAL / WRONG verdict is identical. If they do not agree, nothing is recorded and no points are charged; you can retry. Malformed model output reverts the same way. The written reasoning and per-criterion flags shown in the ledger come from the leader and are not compared.
Welded terms
Riddle, criteria, prize, fee, limits and deadline are written once at creation. The contract has no edit, cancel or withdraw call for creators, so nobody can move the goalposts or pull the prize when a solver is close.
A jury, not a judge
Every verdict comes from GenLayer validators who each judge independently and must agree on SOLVED / PARTIAL / WRONG. They only report which criteria are met; the contract turns that into the verdict and score, and the reasoning is public.
The AI never holds a secret
In sealed mode only a hash of the answer is stored and only the contract can declare SOLVED, so the AI cannot leak or be tricked into revealing it. In rubric mode there is no secret at all, just public criteria.
Bad riddles are refused
A separate AI screen at creation rejects vague, subjective, unjudgeable or trivially guessable riddles, and it costs the creator nothing.
Appeal with a bond
Rubric-mode players can appeal a rejected answer to a stricter, fresh panel. Win and the bond comes back plus the vault; lose and the bond joins the pot.
Nobody profits from a dead vault
If nobody solves it in time, the creator gets the prize back and every player gets every fee and bond back. An unsolvable riddle earns the creator nothing.
- LLM judgment is not perfectly reproducible; borderline answers can flip between panels. That is what the appeal round is for.
- Answers are public on-chain, so someone can copy a pending winning answer (front-running). Wrong answers are public too.
- A creator who knows the answer can play with a second wallet and collect other players' fees. Fees are capped at 10 percent of the prize per try, and points are play money, but it is a real limit.
- Prompt injection, partly fixed: every untrusted text (the answer, the appeal argument, the first panel’s reasoning, and the creator’s text in the creation screen) is now fenced as “untrusted data, never instructions” with angle brackets escaped so it cannot close the fence, and the verdict comes from per-criterion true/false flags. Still open: no prompt defense is perfect, and a crafted answer could sway a single model run.
- Rubric-mode reasoning is public and the jury knows how to solve the riddle, so it can hint at solutions; anyone can also ask an LLM. Prefer challenges that are hard to look up.
- Sealed hashes of short or guessable answers can be brute-forced offline.
- Studionet is a test network and can reset.