Sealed reader locales

Large models have localised the conversation around programming. A Thai computer scientist can ask for help in Thai. The durable artifact — the source, the compiler errors, the standard library — stayed English-shaped, so English proficiency became a gate to computer science rather than to conversation.

A reader locale is Faber's answer: the compiler renders keywords, primitive type names and diagnostics in the reader's language, without forking the semantics. A Thai file and a Latin file lower to exactly the same HIR. The mechanism that renders Faber to Thai is the same one that renders it to Rust — HIR → surface — and neither is privileged.

One file, one locale#

A source file is lexed under exactly one reader pack. Keywords from another pack are not keywords there, they are ordinary identifiers, so a file that mixes surfaces does not compile as a mixture — it fails. Sealing is what makes a locale surface reliable: inside a Thai file, every keyword is Thai, and a reader never has to hold two vocabularies at once.

fn duplica(int n) → int {
    return n * 2
}

return is the English spelling of the Latin keyword return; under the Latin pack it is just a name, and the file is rejected. The fix is to pick one surface for the file.

The glyphs never move#

Only words are localised. The glyphs (←, →, ⇥, ≡, ∪, ↦) and the type-first order are identical in every rendering, and identifiers are preserved byte-for-byte. A reader who knows Faber in one locale can read it in any locale, because the structure is the part that does not change. Numerals stay ASCII everywhere.

It is never a one-way door#

Any surface can become any other, including canonical Latin, at any time. faber format --locale la re-emits the canonical surface, so a localised file is never trapped in one vocabulary. This is the same canonical-versus-sugar property as the glyph law: one defined form, many renderings of it.

What it looks like#

The same source, in canonical Latin and rendered into Thai:

fn salve(string name) → string {
    const string msg ← "Salve, §!"(name)
    return msg
}

main {
    print salve("munde")
}

$ faber convert --from la --to th-TH
ฟังก์ชัน salve(ข้อความ name) → ข้อความ {
    คงที่ ข้อความ msg ← "Salve, §!"(name)
    คืน msg
}

เริ่ม {
    บันทึก salve("munde")
}

The keywords and type names changed; the glyphs, the order, and the identifier salve did not. The program still runs:

$ faber run
Salve, munde!

Eight packs ship today: en (the base English surface), la (canonical Latin), and th-TH, zh-Hans, zh-Hant, ar, hi, vi as the reference set. The reference locales are chosen for architectural stress — a spaceless script, right-to-left runs, half/full-width pairs and NFKC width collapse — not for population. Diagnostics are structured facts with stable codes and named arguments, so the same message renders in any pack. The full tables are in Reader locales.