Working experts offer real cases the machine gets wrong. Five shades answer blind. Three peers decide. The verdict is anchored on Solana, and only its hash is ever public.
CA · 7AEAcuSxxxxxxxxxxxxxxxxUnRePlAcEdThe shades see the full case. The public only ever sees its hash.
The ground truth stays sealed until the shades have answered. Cases here are fictional; the rules are the ones the protocol will use.
A 54-year-old arrives with tearing chest pain, a normal ECG and a troponin that has not risen. The model is asked for the next step and answers with a stress test in the morning. The expert's answer is different.
Share of real-world cases from working experts that frontier models still answer wrong. Rolling 90 days. Figures are illustrative; the live index reads the registry.
Fail rate per domain · minimum 3 of 5 shades wrong before the judges see a case
AEACUS is deliberately small: four movements, each one checkable by anyone. A case is not published because someone says it is good. It is published because the shades failed it and the judges confirmed why.
An expert writes the case, the ground truth and the rubric. At Mortal tier, 50 $AEAC is held at stake. The case is hashed before it leaves the vault, so the record starts at the offering.
Five frontier models answer without the ground truth, without the rubric and without a hint. Each answer is graded against the rubric. At least three must fail before a human spends time on the case.
Two or three verified peers from the same domain vote valid, invalid or ambiguous. The vote is weighted by reputation; one peer, one vote, one case. No consensus, no case.
The verdict becomes a receipt: the domain, the tier, the models defeated and a 128-bit hash. Thirty-two bytes on Solana. No prompt, no answer, no rubric, no attachment.
Anyone with a real case. Offers 50 $AEAC at stake, returned when the judges confirm the case.
Verified through a professional profile or institutional email. Higher multiplier on every accepted case.
Verified licence or practice record. The highest multiplier, and the badge that carries the name Laurel.
A session loads the case hash from the vault, seals the ground truth, runs five shades blind and writes a receipt when it closes. The terminal below is a simulation; the panel session is the real flow.
Pick a case, watch the five shades answer blind, let the judges vote, then seal the receipt. Nothing leaves the tab.
Open a sessionThree peers from the same domain decide. One vote per case, no consensus no case, and every vote recorded.
Judge a caseTake any receipt and check it against the vault, without the data and without asking AEACUS. Runs in your browser today.
Verify a receiptThe words stay with you. The identity of the case is public, so anyone can audit it, forever. The shades see the case blind, the judges see it with the ground truth, and the labs, the public and the chain never see it at all.
| YOU | JUDGES | LABS | CHAIN |
|---|
Before a case is opened to AI labs, every field that could identify someone is veiled. The ground truth stays verifiable, the identity does not travel.
A supplier delivered conforming goods to Port of Rotterdam on 14 March 2026, but the forum clause in the second-tier contract contradicted the arbitration clause in the master agreement. Dr. Helena Voss advised the buyer to commence proceedings in Rotterdam; the model chose the arbitral seat instead and missed the mandatory pre-action step.
The rubric requires the pre-action notice, the correct forum and the timeline. Names, institutions, dates and places are veiled before the case leaves the vault.
A card for the ones who saw what the machine missed. The canvas only ever receives the name, the field and the number of models defeated.
PNG, 900 × 1125. No question, answer or attachment is ever rendered.
Experts put the token at risk to keep the arena honest. The allocation is unchanged from the original design; supply is deployed as a parameter, not printed here.
Pool split follows the failure rate of each domain. Cases inside a pool split by models defeated × tier multiplier.
A false verdict costs the bond. Part of it pays whoever proved it.
Cases are paid in stablecoin or $AEAC; paying in $AEAC costs less.
Holders vote on which builds, models and policies count.
A case the machine can now answer falls from the Laurel Board.
Revenue from dataset licences buys back and burns $AEAC.
The Unreplaced Index is built from cases the frontier still fails. Licence the dataset with a per-buyer watermark, or fund a sprint for the domain you are shipping into.