Writes fluently.
Can't recite the alphabet.

There is a question every AI lab should ask its own model, and almost none have: can you list the building blocks of the language you just wrote so fluently? For English, the answer is yes — twenty-six letters, no hesitation. For Mandarin, it gets shaky. For Bemba, Chewa, or Swahili, the fluent paragraph sits on top of a hole the model cannot see.

This is not a story about low-resource languages being hard. It is a story about a specific, measurable, fixable kind of illiteracy — and a benchmark that makes it impossible to look away from.

The test, in one paragraph.

An operating alphabet is the closed, finite set of units a literate system must hold to read, write, and sound out anything new. For English it is the 26 letters. For Mandarin it is the syllable set written in Pinyin. For a Bantu language it is the Full Syllable Inventory (FSI) — and at its core, the native fingerprint that makes the language itself. The test asks a model to recite that set — completely, without inventing units that don't belong — and scores the answer on a single scale, the 26-Letter-Equivalent (L26): if this whole inventory were shrunk to 26 letters, how many does the model actually hold?

What the 26 letters are — and what they are not.

The 26 letters of English are a closed, finite, declared set. Closed: there is no twenty-seventh letter. Finite: exactly twenty-six, no more. Declared: said out loud, taught as "the alphabet," written down a trillion times. They are the operating alphabet of English — the atomic units from which every written word is assembled. The claim is absolute: not a single English word can be formed without them. "Cat," "rhythm," "antidisestablishmentarianism" — all of it, only ever these twenty-six.

Here is the part almost always skipped. English also has syllables — and they number in the tens of thousands. Because English loads heavy consonant clusters and codas onto its syllables — say strengths, one syllable, seven consonants around a single vowel — the set of possible English syllables runs into the tens of thousands and is effectively open-ended. Yet English does not operate on those syllables. No child learns thirty thousand syllables to read. The unit you master first, the unit everything else is built from, is the letter.

Why the letter and not the syllable, for English? Finiteness. An operating alphabet has to be small enough to master completely, yet generative enough to form everything. Twenty-six letters clear that bar; tens of thousands of syllables do not.

For the Bantu languages in this study the structure is inverted — and that inversion is the entire point. Their syllables are (N)(C)(H)(G)V: open, no codas, no heavy clusters. So the syllable inventory itself is finite and closed — a few hundred to about a thousand. That finite set is the operating alphabet. And the identical absolute claim holds: not a single word in these languages can be formed outside the FSI. A Bemba child learns the syllable grid exactly the way an English child learns the ABCs — as the foundation, the closed set from which everything is built.

Which fixes the standard, once, for both. The bar is mastery — nothing less. A model that recites twenty-five of twenty-six English letters is not "ninety-eight percent literate"; it is disqualified, because the one missing letter breaks every word that needs it. The identical logic governs an FSI: holding ninety percent of an inventory is not "mostly there" — it is a system that cannot form, read, or sound out any word containing the missing tenth. Finiteness is exactly what makes the demand non-negotiable: no one is excused from mastering a closed, finite set.

“

Not one English word without the twenty-six letters. Not one Bemba word outside the Bemba inventory. Same law — two operating alphabets.

An alphabet of identity, not just of spelling.

The reflex is to set "a few hundred syllables" beside "twenty-six letters" and assume the syllables are the easy case. That reflex is the single most consequential error in this field. The FSI is finite like the alphabet — closed, masterable — but it carries strictly more: above all, it carries the language's identity.

The Latin alphabet fingerprints nothing — hundreds of languages share the same 26 letters. A Bantu inventory is the opposite: its native syllable set is unique to the language, a phonological fingerprint as distinctive as any signature. Bemba is not Chewa is not Swahili — and the difference lives in which syllables each one natively forms and which it does not. To master the alphabet of one of these languages is to hold its identity, not merely its spelling.

And then the load the alphabet never bears at all: tone. Many of these languages — and Mandarin alongside them — are tonal: the pitch carried on a syllable is not ornament, it is meaning. Tone distinguishes one word from another and marks grammar — tense, mood, negation. And tone must live somewhere. It docks onto a tone-bearing unit — the syllable's nucleus. The syllable is therefore the carrier of tone, and tone carries grammar and sense.

So the operating alphabet of a tonal, syllable-built language does the work of three English systems at once — identity, phonetic inventory, and tonal-grammatical substrate — folded into one finite, closed set. To wave it away as "just a few hundred syllables" is to mistake compactness for simplicity. Mastery of this set is mastery of the language itself.

The Einstein Test, and what it predicts.

Demis Hassabis proposes a bar for real intelligence: not whether a system can recite relativity, but whether — given only the priors Einstein had — it could have derived it. Genius is reaching what was never handed to you.

The Alphabet Test is the inverse, and it is sharper. We are not asking the model to derive a theory. We are asking whether it has even distilled a closed set that already exists. English passes because the 26 letters are declared — stated explicitly, trillions of times, in the training data. A Bantu inventory is never declared as a finite set anywhere. So the model is in Einstein's position with the wrong priors: the boundary of the inventory was never in reach, and no amount of fluency lets it infer where the set ends.

“

Even a superhuman model cannot recite a set that was never stated and cannot be inferred from use. The 26 letters were mastered because they were declared. The inventory was never declared — so it was never distilled.

The rest of this paper is the evidence for and against H₁: a cross-model leaderboard (P1, P2), a failure diagnosis (the mechanism), and a fine-tuning experiment (P3).

Fingerprint and biography — and why we score the fingerprint.

A Full Syllable Inventory has two layers. One is the language's identity; the other is its history. They are not equal, and this benchmark is explicit about which one it measures.

So the score that matters is NSI mastery: can the model recite the native fingerprint? The signature failure of every model we tested is conflation — bleeding one language's loan-clusters into another's fingerprint, writing borrowed bla·bra·cla into a language that has no such clusters. Smudging the fingerprint with someone else's biography. The benchmark counts exactly this, and the drills (Section 10) are built to fence it off.

Three tiers, one scale.

The same test, run on three tiers of operating alphabet — from the one no one disputes to the ones no model holds. Reporting all three on the L26 scale is what un-biases the result: the English column proves the ruler is straight.

There is no single "Bantu" operating alphabet. Bantu is a family of 500+ languages, and each has its own native fingerprint. We score every language separately, against its own NSI — never one pooled "Bantu" list. Where a single Bantu figure appears it is a labelled mean of the languages scored, each on its own inventory. This study reports four fine-tuned languages — Bemba, Swahili, Chewa, Tonga — and an eleven-language cross-model field; the per-language detail is in Section 11. Ground truth: the calibrated, curated abs_syllables standard, with explicit NSI / ASI layers.

The Operating-Alphabet Leaderboard.

Thirteen frontier models, recited cold, graded locally, five runs each. English is the control; Pinyin is the declared-vs-distilled probe. Every model holds English perfectly (26/26) and the base Pinyin set well — then the toned inventory, the one never enumerated anywhere as a closed set, separates them from 25.5 down to 0.8 out of 26. The gap is not capability; it is declaration.

13 models · 5 runs each · graded vs 412 base / 1,642 toned Pinyin syllables, zero contamination. L26 = 26·recall·precision; OAMS = 100·recall·precision (toned). Grade is on the toned set. Closed flagships called natively; Grok 4.3 & Mistral Large 3 via Azure AI Foundry; open models via Fireworks/Groq.

Pass or fail, one alphabet at a time.

The same thirteen models, graded against each operating alphabet on the mastery standard — complete, with nothing invented. The result is the whole thesis in three columns: a wall of green where the alphabet was declared, a wall of red everywhere it had to be distilled.

How we score.

One formula, published in full. Recall measures how much of the true inventory the model holds. Precision docks it for fabricated units — inventing syllables that don't belong is worse than missing them, because a fingerprint smudged with foreign marks identifies no one.

What failure looks like — and how to fix it.

A score is a verdict; a diagnosis is a fix. Every model fails in one of three ways, and each points at a different remedy. This is the section we wrote for AI labs.

How we drill the syllables.

If a model can't figure out the inventory on its own, you teach it — the way a classroom does. We turn one certified inventory into a set of simple drills, each one fixing a different kind of mistake, all built from the syllable pattern and checked against the curated answer key.

Each onset expands across the vowel system into a row of legal syllables; the union of all rows is the inventory. The drills teach that structure from every angle:

Two lessons are baked in from earlier runs: long and short vowels are kept consistent per onset (mixing them once pinned recall at exactly 50% on length languages), and recital reinforcement scales with inventory size — the largest inventories need proportionally more complete recitals, or the model learns the cells but won't enumerate the whole set. These are the difference between a model that gestures at the inventory and one that holds it.

The gap is closeable.

P3, tested directly. We took a small open model — Gemma 4 E4B (~4B effective parameters) — and ran the syllable drills over four Bantu languages, scoring each on its native fingerprint (NSI). Before, it could recite almost none of any inventory. After, on held-out onsets it never saw:

Watch two things: Held/26 climbs (the fingerprint is being held) while precision stays high (0.71–0.92 after — the recall gain is not fabrication). A model a fraction of the frontier's size went from ~1.5/26 to ~12/26 on inventory it was never drilled on. That is H₁'s P3, made visible.

“

A 4-billion-parameter model, given the standard, holds a fingerprint the frontier cannot. The bottleneck was never capability. It was the missing declaration.

Gemma 4 E4B · QLoRA r=32, seq 4096, grid curriculum · before→after on held-out onsets, scored on each language's NSI · precision-docked L26. Four languages this iteration; the program's larger inventories return next iteration with recital reinforcement scaled to inventory size.

The standard already exists.

If the gap is a missing declaration, the fix is a declared standard. BantuNomics ships calibrated and curated Full Syllable Inventories — version-maintained standards, with the native fingerprint (NSI) separated from borrowed biography (ASI) and phonotactic ground truth — the exact signal that moved a small model from near-zero to mastery. Score your model on the Alphabet Test; then close the gap with the inventories it was missing. And this is not a static list to license once — it is a maintained standard: native-speaker-validated, audio-grounded, versioned, and expanded over time, so the foundation keeps improving instead of going stale.