The known unknown

AI labs have a name for it: jagged intelligence — the admission that frontier models are brilliant in some domains and structurally blind in others, with the blindness usually invisible from the inside until something forces it into view. They call the blindness unknown unknowns.

For the Bantu language family, jagged intelligence manifests in many ways. Tone failure — in a Bantu language, tone is meaning, and orthography never captured it. Morphology failure — the 18-class concord system that runs through every sentence. Vocabulary gaps — the long tail of valid words the writing system never recorded. Code-switching collapse. Audio-text misalignment. The list is long, and each entry sits on top of a deeper layer than the one above it.

That is why we built BantuNomics — and why we are starting with the most foundational of these manifestations. The Full Syllable Inventory. The closed alphabet from which every Bantu word in every Bantu language is composed. Every higher-layer failure listed above sits on top of the syllable layer; fix anything above without fixing this one, and the fix has nothing to anchor on. Fix the foundation, and the higher layers stack up to unblock.

Why the foundation is the lever.

A model that cannot enumerate the syllables of a Bantu language cannot do tone in that language — tone attaches to the syllable. A model that cannot enumerate the syllables cannot do morphology — verb stems and noun-class concord agreement land at syllable boundaries. A model that cannot enumerate the syllables cannot do vocabulary — new words are constructed compositionally from the inventory. A model that cannot enumerate the syllables cannot do code-switching — the boundary between an L1 utterance and an L2 utterance is detectable only when both alphabets are present.

Without the foundation, every higher-layer fix is performed on a missing floor. This is the difference between BantuNomics and an excellent fine-tuning team. A fine-tuner is layering paint on a wall that was never built. We are building the wall. Other teams — Masakhane, Lelapa, Khaya, Awarri, the academic linguists — all do important work above this layer. They need this layer to be solid for their work to stack.

The Bantu family used to be an unknown unknown. It isn’t anymore.

We have named the diagnosis. The orthographic gap that prevents text-trained models from seeing the language they claim to handle has a name now: the Flat Text Problem. The full piece on why Bantu writing was never going to carry the layer that decides meaning lives there.

We have mapped the substrate. The Amina Bucket Standard (ABS) is the schema. BTS-S100 is the FSI document standard. 700+ Full Syllable Inventories, one for every Bantu language — 459 of them released for commercial licensing today — the alphabets of those languages, written down in machine-readable form for the first time in Bantu history.

We have measured the gap. The Bantu Atlas Benchmark scores frontier models against three components rooted in the FSI substrate. Today’s leaderboard: top BABS 702 / 1000. Zero models in Production tier. This is a baby BAB; it will evolve and mature as the ABS schema rolls out. Even in its initial form, it surfaces what every text-built eval misses.

The relation between the syllable and everything else you might want a Bantu model to do is laid out across the proof set: How a Bantu child learns to read (the cognitive evidence), Tone lives in the syllable (why meaning attaches to this layer), FSIs are not data, they’re infrastructure (the architectural argument), and It Takes a Village to Raise AGI (the stakes).

It would be comical if it were academic.

It is not. 400 million Bantu speakers are entering the AI era right now — for healthcare, for education, for livelihood, for identity. The substrate their languages stand on is the substrate every model serving them will be built on. That substrate, until BantuNomics existed, was missing from every frontier system in production.

The substrate is no longer missing. It is now public.

Calling the substrate.

The BantuNomics MCP server runs at /mcp on this host. Public, read-only, no authentication required. Any MCP-aware client can call it today.

list_languages    · search_languages  · get_fsi          · get_syllable
decompose         · compare_inventories · find_loanwords
list_abs_buckets  · score_bab_submission
read_value_proposition · read_mission · get_flat_text_problem
get_bundle_pricing · evaluate_for_ai_lab · request_partnership_call

The REST mirror lives at /api/v1/* for clients without MCP support. Full contract in BTS-API-100.

For Founding Partners, an enhanced MCP is available under subscription credentials — adding audio retrieval (opaque proxy tokens per BTS-API-100 §3.4), multi-component BAB scoring write-side, bulk artifact downloads, and management tools for in-house evaluation pipelines.

We want every frontier model to be better at Bantu, not just one.