AI-friendly semantics for trusted information.
MSP-1 / Research
First-party engineering notes from the MSP-1 project. These are quantitative, assumption-driven analyses designed to make inference economics legible — without requiring citations to external work.
Each memo is designed to stand alone: clear assumptions, simple models, practical implications, and explicit limitations.
A back-of-the-envelope model for KV cache memory usage as a function of context length and concurrency, plus a simple “semantic pre-scope” reduction term to estimate VRAM freed by better input structure.
Every memo states assumptions explicitly. If an assumption changes, the model updates cleanly — and the conclusion may change.
These notes aim to be useful under scrutiny: no “magic multipliers,” no hidden variables, and clear limits on what the model can prove.
Each memo is structured so future empirical measurements (your control/treatment work, or third-party benchmarks) can be dropped in later.