This file is reference context, not a task. It describes the destination so agents always know the end goals. Do not execute these as instructions.
-
Universal base models — Shared foundation consumed by every Token project. One source of truth.
-
Flat, composable — No inheritance hierarchies. Models compose through references. Context handles display, models handle logic.
-
Complete vocabulary — Financial terms (70+), time terms (planned), functions, relationships, maps, styles, identifiers, unifiers, improvement pipeline, agent flows.
-
Language spec — Models become the type system for a compiled language. Layer 0 machine primitives → Layer 1 TokenBase language → Layer 2 generated targets (TS/Swift/Rust/C) → Layer 3 adoption.
-
Bottom-up efficiency — Start from the lowest layer, squeeze the air out, only lose efficiency at each compilation boundary. Don't compress higher-level assumptions into the foundation.
-
AI optimization — Drastically cut consumed resources (tokens, compute, time) to produce results. The language and agent flows should minimize what AI needs to process by giving it precisely structured, pre-refined input. Less waste, same or better output.
-
Living document — New concepts added as they emerge across projects. LANGUAGE.md captures the compiler vision. Insights logged as discovered.
-
Refactor target — After TokenRemote and TokenSports ship, re-abstract everything into TokenBase. Build the project builder framework. Kill busywork.
-
Complete primitive set — The goal is to build any app with just these base models. Everything else is a compound model composed from these primitives. If something can't be expressed as a composition, it's a missing base model — add it. If it can be composed, don't add it.
-
Always look for improvements — Every session, every project, actively look for patterns that should be base models, redundancies that should be collapsed, and missing primitives. TokenBase-first design is the priority — build with the vocabulary even before the package is wired in.