The predictive market @ Polymarket in the 0xPolygon ecosystem has collaborated with Kaito and brevis_ @ brevis_zk to achieve verifiable kaito mindshare markets. This way, users can conduct mindshare prediction transactions for Kaito projects on Polymarket. What is the problem with using a combination of @ eigencloud and @ brevis_zk to implement it? To convince traders on Polymarket that kaito provides data based on a formula (without disclosing the specific formula), there are two aspects that require verifiable calculations: -AI reasoning verifiable: Kaito uses AI reasoning for project mindshare calculations (text calculations on x). If placed in a decentralized environment, the reasoning needs to be verifiable, and EigenCloud reasoning verifiable technology is used to solve this problem -MindShare calculation can be verified: When calculating the MindShare ratio with kaito mindshare, there should be formula algorithms with different dimensional parameters. So how can we prove that kaito is calculated according to the formula without revealing the specific algorithm? This is using @ brevis_zk Let's talk about the verifiable reasoning of @ eigencloud AI here, EigenCloud's method does not directly use zero knowledge proofs (ZK) to validate the entire inference (because LLM inference is computationally intensive and ZK proofs are too costly), but rather achieves it through a combination of inference certainty, re execution, and economic security. The following are the key steps: Making LLM inference deterministic: (Here is the core, EigenAI needs to modify the inference technology stack to restrict some dimensions of randomness in each inference, ensuring the same X input and a definite Z output, such as forcing temperature=0, locking all random seeds, and completely disabling approximation algorithms.) , custom deterministic tokenizer, etc.) Support model: Currently starting from gpt-oss-120B (OpenAI open source model), it will be expanded to more open source LLMs in the future, supporting tool invocation and chain thinking. GenCloud's technology achieves fully deterministic output through fixed seeds, precise control of floating-point operations, and optimization of inference engines. That is, for the same input X and model Y, each run produces exactly the same Z Re execute verification: Inference runs off chain and is executed by operators in TEE or containers The operator provides output Z+signature (proof of execution of specific code). Verification mechanism: Anyone (including challengers) can rerun the same input X+model Y to check if it matches Z. If it does not match, it is evidence of fraud. This is achieved through EigenVerify (EigenCloud's dispute resolution layer): on chain re execution is triggered, the cost is paid by the challenger, but if fraud is proven, the operator's collateral (staked ETH/EIGEN) will be seized Economic and Crypto Security: Restaking based on EigenLayer: Operators use stake ETH/EIGEN as collateral, and the number of stakes is directly proportional to the task risk. Forkability: If a majority of stakers conspire to produce invalid outputs, EIGEN tokens can be forked to reset consensus. Combining TEE proof (hardware level code integrity) and signature to ensure that the execution environment has not been tampered with. The overall operation is under the AVS architecture: AI inference serves as a "service" and is composed of other AVSs such as Oracle and data storage.