Project Thalamus Technical Whitepaper: An Authorship Gate Between Clients and Language Models

Language models compute answers from scratch on every request, generating new variations even when a prompt has been asked before and human reviewers approved a specific response. This technical whitepaper introduces Project Thalamus, a lightweight proxy that acts as an explicit authorship gate between client applications and language models. Thalamus enforces a deterministic operational rule: Approved Prompts: If an approved answer matches via exact string comparison or MiniLM (or other fine-tuned) cosine similarity (>0.72), Thalamus immediately returns the stored text without invoking the model. Unapproved Prompts: If no approved answer exists, the request routes directly to the model every time without exception. Tool-Bearing Requests: Requests containing function definitions or tools always bypass memory and execute through the model to preserve interactive agent capabilities. Version 0.8 details the verified system architecture utilizing independent local JSON files ("brains") rather than external database services, live-tested system endpoints (including OpenAI-compatible completions and MCP Streamable HTTP transport), comparative benchmarks against RAG and traditional caches, and projected impact at scale. All claims in this document have been verified against running code and an automated test suite.