FOR AI LABS AND MODEL OWNERS
Language is the barrier between your model and most of the world.
Stanford researchers found that people who do not speak English are being left behind by large language models, not because the models cannot respond to them, but because the answers are not good enough to rely on. Mātr is the correction layer that closes that gap, on any model, without retraining.
THE REAL CEILING
Fluency got you into the market. Correctness decides whether anyone stays.
A user tries a model in their own language, gets an answer that reads slightly wrong, and quietly stops trusting it. They do not file a bug. They do not complain. They go back to English if they can, and stop using AI if they cannot. Adoption stalls in exactly the markets with the most people in them.
This is not a data problem that the next training run will resolve. Word order, case, agreement, honorifics and information structure fail wherever a language diverges structurally from the training distribution, and they keep failing across model generations.
THE CORE ASSET
The Universal Metalinguistic Framework.
The Universal Metalinguistic Framework (UMF) is a structured inventory of grammatical divergence: 2,003 features across 58 dimensions, describing what each language actually requires rather than what a tokenizer happens to see. It is the reason correction is possible without retraining, and it is the part of this system that cannot be replicated by scaling.
Reranking mechanisms are becoming commodity infrastructure. The framework is not. It is built language by language with expert linguistic input, and it deepens continuously through our Linguistics Excellence Centres.
2,003 features
58 dimensions
2,400+ languages profiled
Fluency scales with data. Correctness does not.
Word order, case, honorifics, evidentiality and information structure fail precisely where a language diverges from the training distribution. These are not edge cases that more tokens resolve. They are architectural mismatches, and they persist across model generations.
In our published evaluation, structural divergence strongly predicted how often a model's output needed correction. A failure that is systematic is a failure that can be fixed at the layer above the model.
FOUR WAYS WE CORRECT
Four methods. Two of them never touch your model.
Different failures need different interventions. The framework underneath is the same in every case.
Candidate reranking
Selects the structurally correct answer from the candidates you already generate.
You sample a small number of outputs instead of one. Mātr scores each against the target language profile and returns the one that fits. Nothing inside the model is touched.
Constraint application
Applies the target language's structural requirements during generation.
The rules a language actually enforces, expressed as constraints the model generates within. Delivered through prompt, logit bias, or validation and repair, depending on what your stack exposes. Nothing inside the model is touched.
Vector bias injection
Corrects recurring structural errors at the representation level.
Where a model makes the same structural mistake in a language repeatedly, the correction can be applied inside the representation rather than after the fact. This is the option for failures that reranking cannot reach.
Typologically weighted synthetic data
Brings low-resource languages to the point where good candidates exist at all.
For languages without the text volume to train on, correction alone is not enough, because the model has nothing correct to select. We generate linguistically complete training data built from the structural profile rather than scraped from whatever happens to be online.
Reranking and constraint application are inference-level. They require no access to your model's interior. The other two are available where you want to go further.
INTEGRATION
One extra call, not a new pipeline.
You keep the completion call you already make. Sample a small number of candidates instead of one, hand them to the correction step, and use its answer in place of the first choice.
Before, standard completion
response = client.chat.completions.create(
model="your-model",
messages=[
{"role": "user", "content": prompt}
],
)
answer = response.choices[0].message.contentAfter, with Mātr correction
response = client.chat.completions.create(
model="your-model",
messages=[
{"role": "user", "content": prompt}
],
n=4, # sample a few candidates
)
corrected = matr.correct(
candidates=[c.message.content for c in response.choices],
target_language="target-language-code",
)
answer = corrected.textValidated results.
up to 2.14x
improvement in structural correctness
83%
average intervention precision
9 languages
in the published proof of concept
Measured on candidate reranking across nine languages. Published February 2026, arXiv 2602.01162.
The head start is the point.
Structural profiles for 2,400+ languages already exist in the system. Deepening one to production quality is a matter of expert linguistic work over about three months, at a small fraction of what per-language data collection and fine tuning costs, and it does not have to be repeated with every model generation. Our coverage strategy targets root languages that anchor families, so each deep profile generalises structural handling to related languages.
BEFORE THE FIX, THE DIAGNOSIS
You cannot correct what you cannot detect.
Existing metrics measure similarity to a reference. None of them can say that an output violates the rules of a language. Mātr identifies the violation first and corrects it second, which means the same system that fixes your output can also measure it. For any provider being asked to demonstrate that a model works correctly in a given language, that measurement is the evidence.
The same system, measuring
Violation detected
Avez vous le livre?
Inversion hyphenation, rule violatedCorrected
Avez-vous le livre ?
Detected first. Corrected second. Scored either way.
Correctness is becoming a procurement requirement.
Regulation and sovereign AI programmes increasingly ask providers to demonstrate that a system works correctly in local languages, not merely that it responds in them. The providers who can show that first will win those markets. The ones who cannot will be asked to prove it anyway.