Can an AI reliably write a language it's barely seen?

Why it matters

It's the fair skeptic's question: an LLM predicts the next token from the huge body of existing-language code, so a language with almost no training data should be hard to write. The answer is structural — Glyph isn't alien. It looks like TypeScript, so the model already knows most of it; the differences are few, enumerable, and each one fails fast at compile time rather than passing silently.

See it

An agent that guesses wrong doesn't ship the guess — it gets a precise, fixable error and corrects in the same loop:

// agent writes, reaching for a TypeScript habit:
if user.admin { grant() }

[E0006] Glyph has no `if`
   ╰─ `match` is the only conditional —
      match user.admin { true => grant(), false => void, }

The error names the Glyph construct and shows the shape, so the fix is one edit away — no source-diving, no guessing. That tight loop is what compensates for the missing training data. In our own adversarial testing, an agent built a real REST API — typed records, tagged-union validation, auth middleware — to a green build in about two iterations. And to remove the last friction, one document (llms.txt) takes an agent from zero to correct, runnable Glyph.