Bantu grammar infrastructure

A model can know every word and still get the sentence wrong.

In Bantu languages, one noun class silently governs the whole sentence: the adjective, the subject and object concord on the verb, the demonstrative, the possessive, the relative, the copula. Miss that agreement chain and the output is fluent nonsense. The Universal Concord Matrix (UCM) turns that chain into an auditable, machine-readable control surface — one language, 31 fixed dimensions, certified.

In plain terms: think of gender agreement in Spanish or French — el vs la, -o vs -a — but far more pervasive. Bantu languages have a dozen-plus noun classes, and the one you choose reshapes the endings on the verb, the adjective, the pronoun, and more, all the way down the sentence. A UCM is the complete, verified table of those matching rules for a language — the answer key a model needs to keep a whole sentence in agreement.

See a live sample Get access / evaluate

The agreement cascade

Why concord breaks text-only models.

Pick a noun class and it cascades outward. Class 2 (people, plural) in Bemba forces ba- on the noun, ba- on the adjective, ba- as the subject concord on the verb, -ba- as the object concord, aba as the demonstrative — and different forms again for the relative, the possessive, and the copula. A model that has memorized vocabulary but not the class system will mix concords from the wrong class and produce agreement errors a fluent speaker catches instantly. The UCM makes that cascade explicit and checkable.

Grammar-aware generation

Constrain adjectives, pronouns, demonstratives, relatives, and verbs to the noun class that governs them.

Translation QA

Detect agreement-chain errors a text-only model can miss even when the vocabulary is correct.

Fine-tuning and evals

Build agreement tasks that test whether a model understands the concord cascade, not just word lookup.

ASR/TTS language support

Connect surface forms, class systems, and morphology for pronunciation-aware language products.

Dictionary expansion

Attach every noun and term to the agreement system it activates across a sentence.

RAG and tooling

Retrieve the right concord row as a control surface during generation or validation.

The contract

31 canonical dimensions, grouped.

Every language is mapped across the same fixed 31 columns, so a class row reads as a complete agreement profile and languages stay comparable without being flattened into each other.

Canonical prototype Source-certified Comprehensive structural Draft

Class Identity

Class_ID Class Identity

Number Class Identity

Semantic_Profile Class Identity

Is_Honorific Class Identity

Nominal

Noun_Augment Nominal

NP_Base_Prefix Nominal

Pre_Vocalic_Form Nominal

Adjectival

Adjective_Augment Adjectival

Adjective_Agreement Adjectival

Verbal

SP_Affirmative Verbal

SP_Negative Verbal

Object_Prefix Verbal

Tone

Default_Tone Tone

Pronoun

Absolute_Pronoun Pronoun

Copula_Affirmative Pronoun

Copula_Negative Pronoun

Emphatic_Alone Pronoun

Emphatic_Self Pronoun

Associative_Pronoun Pronoun

Demonstrative

Demonstrative_Immediate Demonstrative

Demonstrative_Proximal Demonstrative

Demonstrative_Medial Demonstrative

Demonstrative_Distal Demonstrative

Demonstrative_Negative Demonstrative

Demonstrative_Emphatic Demonstrative

Demonstrative_Contrastive Demonstrative

Connective

Associative_Prep Connective

Relative

Subject_Relative Relative

Object_Relative Relative

Relative_Negative_Marker Relative

Nominalizer

Ntu_Nominalizer Nominalizer

Coverage

19 languages published.

Names and certification levels are public. The full matrices, cross-language comparison, and machine-readable exports are available under evaluation and license.

Taste the Bemba sample

bem Canonical prototype

Bemba (IciBemba)

Bemba · M42 · Zambia

Level 3 target

View public sample

lun Comprehensive structural