POKEBALL: Protocol for On-chain Knowledge Encoding Blockchain AI Latent Loader

We present POKEBALL, the first demonstration of a reconstructable large language model (LLM) weight representation stored permanently in a smart contract on a public blockchain. A hypernetwork is trained to compress the full tensor weight distribution of TinyLlama-1.1B (Q4_K_M quantization, 637 MB) into a 1,024-byte latent seed vector z in R^256, achieving a compression ratio of 654,296x. The seed is inscribed as immutable contract storage on the Aleph Zero blockchain via an ink! smart contract, queryable by any on-chain or off-chain agent. Given the open-source hypernetwork architecture and the on-chain seed, the full approximate weight distribution of TinyLlama-1.1B is deterministically reconstructable. This work introduces a new paradigm for AI model permanence: model provenance, integrity, and approximate reconstruction capability, encoded in a few hundred bytes, living forever on a decentralized ledger. We describe the hypernetwork architecture, the compression protocol, the on-chain data schema, and the smart contract ABI, and discuss implications for decentralized AI, model authenticity, and browser-native LLM inference via Transformers.js.