The writing is not incomplete — it is efficient

In most everyday writing across Bantu languages, standard orthography generally avoids diacritics, tone marks, and double vowels, except in academic or linguistic contexts. This is efficient for in-group communication because native speakers can infer the intended pronunciation, tone, and meaning from context.

The challenge is that this creates a flat text problem: the written form hides important oral features of the language. Tone, vowel length, stress, and syllable-level distinctions may be obvious to native speakers, but they are not visible in the text itself. As a result, the same writing system that works well for fluent speakers can become incomplete or ambiguous for outsiders, learners, researchers, and language technologies that do not already possess native intuition.

A frontier AI model is simply the newest outsider in that line — and the one with the least excuse. It has the text, and only the text: the half the writing system deliberately left for the reader to supply.

One spelling, ten meanings — hear it

Take the Bemba word ulebomba. Eight letters. In standard writing it is a single, flat string. To a Bemba speaker it is up to ten different utterances — the pitch contour, the vowel length, the phrase-edge all change the word or the grammar. To a model trained on text, all ten are the same string. Press play on each and hear what the page cannot show.

ulebombau-le-bo-mba You are working — present statement A plain declarative contour resolves the statement reading.
ulebombau-le-boo-mba You are getting wet — present statement A length-and-tone shift sends the same spelling to a different word.
ulebombau-le-bo-mba You should work — modal statement The grammatical force is in the sound layer, not the letters.
ulebombau-le-bo-mba Are you working? — yes/no question The phrase-boundary edge turns the reading into a question.
ulebombau-le-bo-mba The one who is working — relative reading The utterance role changes while the standard spelling stays put.
ulebombau-le-bo-mba You are really working — emphatic reading Emphasis, timing and voice quality live outside the flat string.
ulebombau-le-bo-mba The one who is working? — question with focus Focus and boundary cues separate this from the statement reading.
ulebombau-le-boo-mba The one who is getting wet? — question + lexical shift Lexical meaning and sentence force both depend on the acoustic layer.
ulebombau-le-boo-mba Are you getting wet? — direct question The same written form becomes a direct question in speech.
ulebombau-le-boo-mba You should be getting wet — modal progressive Ten meanings, one flat spelling — the collision is complete.

Ten meanings. One flat spelling. Every distinction that separates them lives in the sound, not on the page.

The layers the writing drops

None of these survive into ordinary written Bantu; all of them are required to decide meaning. This is the layer a text-only model was never given:

L1 · Mora

Vowel length

A long nucleus that ordinary spelling does not mark — often the difference between two distinct words.

L2 · High / low pitch

Tone

Pitch on the vowel nucleus. The syllable is the tone-bearing unit; same letters, different pitch, different word.

L3 · Lowered high tone

Downstep

A high tone realised lower than expected because of tonal structure — it can resolve relative vs declarative.

L4 · L% / H% edge

Phrase boundary

The pitch at a phrase edge separates statements, questions and discourse boundaries that flat text hides.

L5 · Phonation cue

Voice quality

Breathiness, creak and related cues carry emphasis and emotion — lost entirely on the page.

Every one of them docks onto the syllable — the tone-bearing unit. So the missing meaning-layer is anchored to the very unit a model cannot enumerate. A reader puts the layers back from native knowledge; a text-trained model has nothing to put them back from.

More text will not fix this

The reflex is to scale the corpus. But every additional Bantu page on the web has the same layer missing. Adding more flat text adds more of the same omission — none of what is absent. The information is not sparse in the text; it is absent by construction. You cannot recover a dimension the writing never encoded by collecting more of the writing that omits it.

And it is invisible to the usual measurements, because text-based evaluations are also flat text — a fluent-but-tonally-wrong output scores as correct on the page. The failure only exists in the air, or against native ground truth. That is what makes it a true unknown unknown: not just unsolved, but unmeasurable with text-only tools.

The fix: supply the intuition the writing assumes

If the flatness is the writing assuming native competence rather than lacking a feature, then the fix is not to repair the orthography. The fix is to supply the native competence the writing takes for granted — in a form a machine can hold.

That is exactly what BantuNomics provides, in two interlocking layers: the Full Syllable Inventory — the closed set of a language's legal syllables, the bearers tone and length attach to — and the consented native audio that gives those bearers their spoken values. The inventory says which syllables exist; the audio says what each one actually sounds like. Together they make flat text decidable again. And it is recoverable precisely because, as native writing shows, these distinctions are real and markable — just conventionally left out.

Prove the gap on your model Read: it takes a village to raise an AGI Why the FSI is infrastructure Initiate Partnership