Exact Integer Output Projection and Transaction Packing for a Small Solana Decoder

This report describes a bounded systems optimization in SEA LLM, a small decoder implemented as a sequence of Solana program instructions. The inspected implementation has width 64, eight transformer layers, a 50,257 entry vocabulary, and a maximum context of 32 tokens. Its output projection uses signed eight bit operands. Expanding the fixed width dot product preserves its exact integer result while removing loop and indexing overhead from the compiled SBF workload. Project documentation reports a reduction from approximately 1.215 million to 0.426 million compute units per vocabulary chunk and from approximately 162.9 million to 112.4 million compute units for a complete first token. A historical deterministic fixture also reports packing 382 instructions for a one token prompt and one cached decode into 235 transactions, with a peak of 1,276,710 compute units per transaction. These figures are existing project reports, not measurements newly reproduced for this manuscript. We inspect the kernel, schedule, benchmark, and client planner to explain their contracts and the limits of the evidence. The result is a component study of exact arithmetic and transaction organization, without a claim of trained model quality, live network throughput, or large model inference. Preparation of this report was assisted by Codex. The research programme is directed toward human agency, financial inclusion, and evidence-based accountability. Intended social benefits are distinguished from measured technical results in the manuscript. The manuscript and newly authored analysis are offered under CC BY 4.0. Any supplied source files retain the license terms stated in the supplement; the record license does not replace third-party source licenses.