Languages

Covering "most of the 54" is not 54 translations. Four languages reach the official language of about fifty countries; the work that actually matters is the second tier, where people count money in the language they think in.

Three tiers

Tier 1 — pivot languages. English, French, Arabic, Portuguese. Between them these are official or co-official in roughly fifty of the fifty-four member states. Everything ships in all four or it does not ship. Swahili belongs here in practice rather than by status: it is a working lingua franca across the East African Community and reaches more daily speakers than Portuguese does.

Tier 2 — the languages people actually transact in. Ordered by speakers, not by state:

Hausa       Nigeria, Niger, Ghana, Chad          ~80M
Amharic     Ethiopia                             ~35M
Oromo       Ethiopia, Kenya                      ~37M
Yoruba      Nigeria, Benin, Togo                 ~45M
Igbo        Nigeria                              ~30M
Zulu        South Africa                         ~28M
Somali      Somalia, Ethiopia, Kenya, Djibouti   ~22M
Fula        Sahel, fifteen countries             ~35M
Akan/Twi    Ghana                                ~20M
Shona       Zimbabwe                             ~15M
Malagasy    Madagascar                           ~25M
Kinyarwanda Rwanda, Burundi                      ~20M
Wolof       Senegal, Gambia, Mauritania          ~12M
Bambara     Mali                                 ~15M
Xhosa       South Africa                         ~11M

Tier 3 — everything else. Real, but reached through community contribution rather than commissioned translation. The system has to make that possible from day one or it never happens.

Note what tier 2 does that tier 1 cannot: Fula follows the Sahel across fifteen borders, and Hausa crosses four. These are trade languages. A cross-border payments network that speaks only the coloniser's language at the border post has missed who is actually moving money.

The hard parts are not translation

Arabic is right-to-left. This is a layout problem, and it is the single largest piece of engineering here. Every flex direction, every icon that means "forward", every transaction row with an amount on one side and a name on the other has to mirror. CSS logical properties (margin-inline-start, not margin-left) make this nearly free if adopted before the UI is built and extremely expensive afterwards. Adopt them now, while there are four frontends and not forty screens.

Amharic and Tigrinya use Ge'ez script; Tifinagh and N'Ko are not Latin either. These fonts are absent from many Android devices in the field. Either subset and self-host the fonts or the user sees boxes — and a payment screen full of boxes is worse than English.

Text expands. French runs 15–25% longer than English, German more, and compound-forming languages worse still. Any button sized to its English label breaks. Test the layout in the longest language, not the shortest.

Arabic has six plural forms; most Bantu languages have noun-class agreement that plural rules do not model at all. Use ICU MessageFormat and let the catalogue carry the rule, rather than concatenating a count with a noun in code.

Numbers and money are locale data, not string data. Decimal separators, digit grouping, and currency symbol placement all vary — and Arabic may use Eastern Arabic numerals. Format through Intl.NumberFormat with the currency and locale, never by string manipulation.

Never translate the chain's errors

The chain returns errors in English from Go, and they are not user-facing text — they are diagnostics that happen to be in words. Translating them means shipping a translation pipeline that runs at the speed of chain upgrades and produces sentences no user asked for.

Instead: the chain returns stable error codes; the client owns the message. ErrDuplicateRef becomes a catalogue key, and the catalogue says, in Wolof, what the user should do about it. This also means an error's wording can be fixed without a consensus-breaking upgrade, which is the same reasoning that keeps search and presentation out of consensus.

This has a prerequisite worth naming: every module's errors need stable, enumerated codes that clients can switch on. That is worth auditing before translation starts, not after.

Money words are not ordinary words

"Balance", "pending", "fee", "settled", "frozen", "available" carry meaning a central bank will hold you to. A volunteer translator rendering "pending" as something closer to "maybe" produces a legally different product.

So the glossary is a deliverable, not a by-product: a fixed list of financial terms with a definition in English, translated once by someone accountable, and locked. Community contribution applies to the rest of the interface, not to these. x/enforcement freezing an account and the word for "frozen" have to mean the same thing in every language, because someone will read it in court.

Architecture

One catalogue, shared by every surface. The four frontends, the site, and the docs all consume the same keys from the SDK — a term translated once is translated everywhere, and a term that drifts is visible immediately.

ICU MessageFormat, for plural and gender rules the catalogue can express and code cannot.
Locale resolution: explicit user choice, then device locale, then the country's tier 1 language, then English. Never infer from IP.
Lazy-load catalogues. Fifty locales in one bundle is a mobile data bill in a region where data is the constraint.
Language is a device preference, not account state. It does not belong on-chain: it is not consensus, and putting it there makes changing your interface language a transaction that costs money.

Sequencing

CSS logical properties and an RTL audit, before more UI is built. This is the decision that gets expensive with delay.
Extract strings to a catalogue with ICU, English as source.
Stable error codes across every module, replacing English strings at the client boundary.
Financial glossary, locked, translated accountably.
Tier 1 — French, Arabic, Portuguese, Swahili.
Community contribution, with the glossary enforced and the tier 2 list as the target.

Step 1 is the only one that is harder tomorrow than today.