Programming languages were made for people typing lines of text. AI reads structure, carries intent and checks its own work. EML is a family of languages built for both readers, five of them public: EML-U evolves existing languages, EML-P runs today, EML-NOVA builds programs as typed graphs, NoGlyph separates a program from its visible text, and CVSG writes programs in Chinese over a semantic graph.
Each line asks a different question about programs: how a language itself can evolve, how a program runs today, how an AI builds one as structure, what a program is without its text, and how it can be written in Chinese.
A language need not be rebuilt from scratch to gain new capabilities. EML-U attaches explicit semantics to existing languages and composes them: one semantic model, projected into C++20, Rust 2021 and EML-P/Python, with evidence for every supported transformation.
The stable, linear, low-ambiguity subset of EML-U. It transpiles deterministically to Python, runs in your browser, and every published case is checked against real Python.
To an AI, a program is a typed graph with a stable identity. EML-NOVA builds, changes and verifies programs as structure; text is one projection among several.
A program is not its visible text. In NoGlyph the visible source can be completely blank: the program is a canonical set of constraints, which a deterministic resolver turns into a program graph and then into C for a conventional compiler.
Chinese is how a program is written, not what it is. The canonical program is a typed semantic graph whose states take values on a spectrum, and a versioned rulebook decides how each sentence is read. Rules run as deterministic, replayable world-state transitions.
There is one more line in this family. It stays private for now.
One small intent, followed through three members of the family. The EML-P program is real: it is transpiled and executed every time this site is built, and its output matches CPython byte for byte. The EML-NOVA graph is real too: the standard library’s own, verified.
“Add up the squares of 1 to 100.”
Semantic attachments on an EML-P core, sketched in words. EML-U’s implementation attaches and composes semantics as structured records; its final surface notation is not fixed yet.
N^+100 Σ(i^2, i in [1:N]) => r r^0
N = 100 r = sum(i**2 for i in range(1, N+1)) print(r)
The same intent in NOVA’s standard library: sum_of_squares_to, drawn from its own graph (n is the length of its input). Verified like every library function: equal to n(n + 1)(2n + 1)/6 in exact arithmetic, on 40 test cases.
Open the functionLanguages were designed for people typing linear text. AI reads structure, keeps context and checks its own work, so meaning, structure and verification belong inside the language, not around it. In EML-NOVA, NoGlyph and CVSG the program is structure, and text is one way of looking at it.
EML-P is a subset of EML-U. Whatever cannot be carried down to EML-P is kept as metadata or marked unsupported. The family’s rule is that nothing is dropped silently.
Every EML-P case on this site is transpiled, executed and checked against real Python before it is published. There is no LLM in the core chain: AI may suggest, the toolchain decides.
People read these pages. Agents read the same site through a public machine layer: llms.txt, a manifest, the specifications and tools they can call.