Agents write it. Humans read it. English is the high-probability default; the locale is theirs; the meaning is one program.
The same HIR then projects onto measured compilation targets. Application lowers currently include Rust, TypeScript, Go and Swift; the MIR lane includes LLVM IR and WebAssembly; the GPU lane names Metal and CUDA. Support is stated target by target. Bounded dual-backend training on Metal and CUDA is the current proven device claim; end-to-end device GPU inference is not shipped.
## What it looks like.
main {
const tensor<f32, [2, 3]> a ← empty
const tensor<f32, [3, 4]> b ← empty
const tensor<f32, [2, 4]> product ← a · b
print product
}
English reader surface. Those keywords were chosen for model probability, not human fashion. The same HIR renders into Thai or Chinese; a human reads their keywords, an agent emits the sealed pack, the matmul glyph and the names stay put.
Declarations are type-first (`const tensor<f32, [2, 3]> a`, not `a: Tensor`). Runtime binds use `←`. Tensor work that is a nested index walk in ordinary code is a glyph or a method: `·` matmul, `⊗` outer product, `⊙` Hadamard, `↦` convert, `.mean()` / `.sum()` for reductions. Locales are sealed: `fn` and `functio` do not mix in one file.
## Distinctive moves.
- **Agents author; humans read.** The surface is mechanical so a model can emit it. Glyphs are for the agent and the reader, not for a human to hunt on a keyboard.
- **English keywords chosen for model probability.** Latin is canonical interchange. The project states the English locale was set keyword-by-keyword by a council of different models for the most stable, highest-probability hit on each concept.
- **Native-language loop through HIR.** Speak to the model in Chinese or Thai, write and read the `.fab` in that locale, compile the same program.
- **Sealed reader packs.** One locale per file. English is the default writing surface; Latin is the interchange template, not a required writing language.
- **Predictable tensor surface.** Rank-aware operators carry shape contracts into typechecking (inner-dimension unification on `·`, identical-shape on `⊙`).
- **Agent-facing docs shipped with the product.** `faberlang.dev/llms.txt`, a skills catalogue, and `faber explain` for glyphs, keywords and diagnostic codes.
- **Honest capability ladder.** Reader-localised source is shipped. Bounded Metal/CUDA training is proven. Device GPU inference and multi-device execution are not current product claims. Version 1 is experimental; stability is a version 2 claim.
## Maturity.
A working `faber` CLI ships as a tagged release archive; v1.0.0 publishes a macOS arm64 build. The language, public libraries (including Gradus, the MIT autograd and ML library), examples and the documentation site are MIT. Radix, the compiler, is closed while it is under active development; that is presented as temporary, not as a permanent fence. Public GitHub organisation `faberlang` holds the open surface; the Haskell project at `faber-lang/faber` is unrelated.
The strain under real use is the usual early-language one, plus a specific honesty gap: the tensor and inference *surface* is large (Gradus is a substantial Faber library) while device residency and GPU execution remain compiler and host concerns. A catalogue reader should not take the glyph surface as a shipped serving product.
## Agent tooling.
`https://faberlang.dev/llms.txt` is the machine index; `llms-full.txt` is the expanded map. The site content-negotiates an agent guide at `/en-US/` and publishes skills for install, language, packages, examples and corpus. `faber explain` answers glyphs, keywords, grammar terms and diagnostic codes. The inclusion claim is the design intent — agents as authors of a locale-rendered, mechanically predictable HIR, with an English keyword surface chosen for model probability — plus that shipped tooling, not a runtime chatbot wrapped around another language.
## Design DNA
- **[Tacit](https://agentlanguages.dev/languages/tacit.md)** *(Syntactic)* — Tacit also treats the tree as more authoritative than one national-language spelling. Faber uses that split so a person can author with a model in their own language; Tacit uses it to canonicalise a single text and drop names toward De Bruijn indices.
- **[X07](https://agentlanguages.dev/languages/x07.md)** *(Syntactic)* — X07 deletes text and edits JSON ASTs. Faber keeps text on purpose: the agent still writes a .fab file, just a sealed, locale-rendered one a human can read.
- **[Zero](https://agentlanguages.dev/languages/zero.md)** *(Verification)* — Both ship agent-facing compiler output (skills, structured explainers). Zero's bet is verification and typed repair; Faber's is a predictable authoring surface plus native-language reader packs.
- **[Vera](https://agentlanguages.dev/languages/vera.md)** *(Verification)* — Both publish llms.txt and skills. Vera removes parameter names so the model cannot lean on them; Faber keeps readable names and moves the language barrier into locale packs so the model can write in the human's language.
## Timeline
- **Jul 2026** — The public