MĀTR,THE UNIVERSAL INTERPRETATION MODEL
What sounds right
is not always right.
Every language has structure. Mātr makes AI follow it.
Models predict plausible words, not correct grammar. They learn the shape of a few languages and apply it to the rest, creating a critical gap in linguistic precision.
Drag to spin. Tap any language.
The gap is not data. It is structure.
01
Grammar is not universal.
Word order, case, honorifics. What a sentence needs depends on the language it is in.
EnglishSVO
SinhalaSOV
ArabicVSO
Same sentence, three required orders. Among documented languages, verb-final order is the most common. The text most models learn from is overwhelmingly SVO.
02
Fluent is not correct.
A model can sound natural and still be wrong. In a hospital or a courtroom, that is the expensive kind of wrong.
Model output
She gave the doctor his results yesterday mornings.
Structurally correct
She gave the doctor her results yesterday morning.
Staged in English so you can see it. This is what AI output looks like in most other languages, every day.
03
Scale does not fix it.
More parameters make a model more fluent everywhere. More correct only where the structure already matched.
Read this out loud
I have been knowing her since Wednesday.
Every word is English. The grammar is almost right. Something in you refuses it anyway.
THE COVERAGE
Every dot is a living language.
7,000+ living languages. About 100 have the text volume to train a model properly. For the rest there is neither enough data nor a deep enough grammatical description to manufacture it. So we describe the structure directly. 2,400+ so far.
about 100
Meaningfully covered by major AI models
about 200
Largest open translation efforts
2,400+
Structurally profiled by Mātr
the remainder
No meaningful AI support
Where Mātr fits.
Mātr is a Universal Interpretation Model. It corrects the grammar a language actually requires.
It reads from the Universal Metalinguistic Framework, our inventory of how languages diverge.
Output
She gave the doctor her results yesterday morning.
Model agnostic. Inference time. No retraining.
- 01Your data goes in.
- 02The model produces something fluent.
- 03Mātr corrects what the language actually requires.
- 04The reader gets something they can trust.
How it works.
Profile
Every language described across 2,003 features in 58 dimensions.
Measure
How far this language sits from the structure the model assumes.
Correct
At inference time, the structurally correct output is selected.
Live correction
See it happen.
One sentence, two answers, and the grammar that separates them.
English source
The children play in the garden.
Target language
සිංහල Sinhala
The model
ළමයි ක්රීඩා කරනවා උද්යානය තුළ.
Mātr
ළමයි උද්යානයේ සෙල්ලම් කරයි.
Word order
English puts the verb in the middle. This language puts it last, so the sentence was reordered.
Locative case
"In the garden" needs a word ending that marks place. That ending was added instead of a loose preposition.
Sense selection
The word for play has several senses. The one about children having fun was chosen over playing music.
Illustrative examples. The production engine returns the same three parts for any supported language.
Measured, not asserted
Published evaluation. Independently checkable.
- 2.14x
- improvement in structural correctness, at best
- 83%
- average intervention precision
- 9
- languages, across 340+ evaluated cases
Standard translation metrics scored the corrected and uncorrected outputs as equivalent. Native speakers did not. Read the method and the results.
The film
Why AI fails at language, and what interpretation changes.
FOUR WAYS IN
Who this is for.
HOW THE FRAMEWORK GROWS
Every profile makes the next one better.
The Universal Metalinguistic Framework deepens through the Linguistics Excellence Centres. Describe one language, and every language that shares that pattern gets better.
- 2,003
- typological featuresstill being extended
- 58
- dimensionsstill growing
- 2,492
- languages profileddeepening every week
The largest public typological record describes languages across roughly 195 features. The framework is an order of magnitude deeper than the public typological record, and it keeps deepening.