RESEARCH
We publish what we validate.
Structural correctness is measurable. That is the whole basis of the system, so we hold ourselves to showing it.
THE FRAMEWORK
The Universal Metalinguistic Framework.
The Universal Metalinguistic Framework (UMF) is a typological inventory of 2,003 features organised under 58 dimensions, describing the ways in which languages structurally diverge. Every language in the system is represented as a structured profile within it. Divergence between any two profiles is computed inside the engine and drives correction at inference time.
The divergence is not a niche concern. Anthropic’s own research reports that Claude behaves differently across the twenty languages it measured, in ways the lab cannot yet trace to its training data.
The framework’s contents are described here. The computation over them is proprietary and is not published.
2,003 typological features
58 dimensions
2,400+ languages profiled
PUBLISHED
A Universal Metalinguistic Framework for Inference Time Correction of Structural Divergence in Multilingual Language Model Output
arXiv 2602.01162, February 2026
A proof of concept across nine typologically diverse languages, using UMF guided candidate reranking over standard model output. Structural correctness improved by up to 2.14x against baseline. Average intervention precision was 83 percent, meaning most corrections the system made were the right ones. Structural divergence correlated with correction frequency at r equals 0.82, which is the finding the rest of the platform rests on.
This preprint covers candidate reranking, the first of the methods we have built. The work below extends beyond it.
Profile, measure, correct.
Every language is described as a structured set of typological features. The divergence between any two languages is computed inside the engine. At inference time, candidate outputs are scored against the target profile and the structurally correct one is selected.
Source
The children play in the garden
Baseline output
ළමයි උයනේ සෙල්ලම් කරනවා
What changed
- Word order enforced SOV
- Locative case applied to the garden
- Recreational sense of play selected
The divergence computation itself is proprietary and is not published.
WHAT WE HAVE NOT PROVEN YET
Honest limits.
- 1. The published proof of concept covers nine typologically diverse languages. We have not yet shown the same gains hold across the full 2,400+ languages profiled in the system.
- 2. Our evaluation measured structural correctness and intervention precision. We have not separately measured downstream effects on fluency, tone, or user preference at scale.
- 3. The correction method has been validated against standard model output at inference time. We have not yet validated it against locally fine tuned or open weight models in production.
- 4. Structural divergence correlating with correction frequency at r equals 0.82 is a strong result, not a proof of causation across every feature in the framework.
THE WIDER WORK
Four correction methods and a tokenization approach.
The published proof of concept covers one of them. Papers on the following are in preparation.
Candidate reranking
Selecting the structurally correct output from candidates the model already produces.
Published, February 2026
Constraint application
Applying a language's structural requirements during generation rather than after it.
In preparation
Vector bias injection
Correcting recurring structural errors at the representation level.
In preparation
Typologically weighted synthetic data
Generating linguistically complete training data from the structural profile, for languages without the text volume to train on.
In preparation
UMF-aware tokenization
A tokenization approach developed here, designed around morphological structure rather than frequency in an English-dominant corpus.
In preparation
The framework underneath all five is the same. Each method addresses a different class of failure.
MEASUREMENT
Detection is half the system.
Established metrics compare an output to a reference and report similarity. None of them can state that an output violates the grammatical rules of the language it is written in. Because the framework describes what a language requires, it can identify the violation directly, which makes the same system usable as a measurement instrument as well as a correction layer.
Reference similarity
One output right, one wrong, both scored alike.
The framework
The violated rule, identified and named.
Standing on public typology.
The baseline profiles in the Mātr system are built on decades of open typological scholarship. We are grateful for it and we cite it.
Language diversity statistics used across this site are drawn from these sources and from Ethnologue.