Arjuna Badger Press

Buabantu — the Real-Language Router API

Part of the technology exposé. Buabantu is the language arm of the studio, promoted out of the press's own translation engine into a standalone product. Closed beta — ABP account required.

Buabantu is OpenRouter, but for register and dialect. You send text and a target — a language, how colloquial to be, who the audience is — and get the meaning back in the language people actually speak, not textbook flatness. Arjuna Badger Press is just one consumer of the API, not its owner.

It began as the press's internal People's Language translation feature (its own page) and grew into something a company can call directly.

The shape: a router with a value-added layer

The base — neutral routing — is a commodity. If a frontier model, or a specialist like Lelapa or GhanaNLP, does a given slice best, Buabantu routes to them and pays their cost. That is the OpenRouter ethos taken seriously: it routes to the big engines, it doesn't try to defeat them.

The value — and the whole product — is the VAS bundle layered on every byte. The governing rule is never a dumb pipe: nothing passes through untreated.

flowchart LR
    DEV[Developer app] -->|bua_live_ key| GATE
    subgraph BETA[Buabantu beta · /api/buabantu/*]
      GATE[Auth + rate-limit + quota<br/>ABP account mandatory]
      subgraph VAS[The VAS bundle — never a dumb pipe]
        COR[Corpus-first correction<br/>human fixes weight 100]
        REG[Register dial · temp 0→1<br/>inferred on inbound]
        FAI[Faithfulness rules<br/>names · notes · length-ratio]
        POL[Police + Judge guardrail<br/>regex · small model · fail-closed]
      end
    end
    GATE --> COR --> REG --> FAI --> POL --> ROUTE
    subgraph ROUTE[Neutral routing — pay the best engine gladly]
      OR[(OpenRouter aggregator)]
      PARTNERS[Lelapa · GhanaNLP · frontier models]
    end
    POL --> OUT[meaning + sentiment / localized text]
    ABP[Arjuna Badger Press] -.just another consumer.-> GATE

VAS layer	What it adds over raw routing
Corpus-first correction	Human fixes (weight 100) overlay any engine's raw output — the proprietary ground truth.
Register dial	`temp` 0 (formal) → 1 (street), tuned to the audience's age/demographic.
Faithfulness rules	Proper names preserved, no translator's notes, per-language-family length-ratio checks.
Police + Judge guardrail	Two-layer safety on both directions — see below.

It runs both directions

Outbound (corp → street): turn corporate-speak into the language a market actually uses — marketing, campaigns, support replies that don't read like a memo.

Inbound (street → corp): the higher-value half. Decode what people say — slang, code-switching, dialect — into meaning and sentiment the business understands. The twist: register is inferred from the text and used to resolve meaning, because the same word inverts by speaker. "This is the shit" from a low-formality teen is strongly positive; from a high-formality elder, negative. Get that wrong and an inbound system reports sentiment exactly backwards.

The faithfulness scan that taught the engine a rule

The first real run of the faithfulness VAS — over the four shipped RESONANCE translation editions (Afrikaans · French · Spanish · isiZulu) — proves the never-a-dumb-pipe posture in miniature.

Names preserved, no note-leakage, every chapter aligned — clean. But the length-ratio check flagged isiZulu at 0.66 (a third "shorter" than English). On inspection the shortfall was flat across every chapter — the signature of a language property, not an omission. isiZulu is agglutinative: it folds articles, prepositions, and pronouns into single inflected words, so a faithful Zulu translation genuinely carries ~30–35% fewer whitespace words.

The flag was a false alarm — and catching it was the point. The lesson folded back into the spec: a length check needs per-language-family baselines (Bantu ≈0.6–0.7, European ≈0.9–1.15) or it cries wolf on every Nguni/Sotho edition forever. The gate is allowed to be wrong; it is not allowed to be wrong quietly.

The beta is real, and gated

A working API is mounted today: POST /api/buabantu/translate, POST /api/buabantu/decode, key management, GET /api/buabantu/status, and a landing page at /buabantu.

An Arjuna Badger Press account is mandatory before any key — minting one needs a verified ABP login,

enforced server-side. No path to a key without an account.

Keys are sha256-hashed at rest (shown once, bualive… prefix), rate-limited, and

monthly-quota-capped — the "limited" in closed beta.

Built on the press's own real_language.py engine and the same police/judge guardrails the studio

uses internally (saas/guardrails.py + judge_client.py) — reused, never rebuilt.

Why it's a fair fight

There are real efforts in African-language AI (Lelapa, GhanaNLP, Masakhane) — but no clear commercial winner, held back by the compute "infrastructure ceiling" and data scarcity. Buabantu doesn't try to out-compute anyone. It competes on the angle: register as the core dial, inbound decode with register-conditioned sense, and corpus-first human correction that out-datas the size race. Rollout follows corpus depth — South Africa first, then Swahili, then the rest of Africa as each language clears review. (A trademark clearance on the name is still owed before commercial launch.)

← Back to the technology exposé · related: People's Language · the police + judge guardrail

Craft Library · Place Wiki · About the press · View this document on GitHub · Write with us

← Back to the library