She didn’t call.
typeassertion
polaritynegative
statusasserted
≈ 0.5one vector
I propose to consider the question, ‘Can machines think?’
Rheme reads what you mean.
A deterministic semantic runtime that acts like a language server for meaning. It resolves prose in nanoseconds, and never flattens what was said.
what it sees
Ten ways to read a sentence.
Rheme reads ten axes of meaning in a single pass. It can tell a claim from a quote and certainty from a hedge.
According to Dr. Reyes, the runtime can resolve attribution in every test, a result she calls genuinely remarkable, and honestly, it never flattens what was said, because meaning, in the strict sense, survives the parse. Plainly put, it works.
play
Theme and Rheme
Every sentence has a topic or theme (what we’re talking about) and a rheme, which is what you're saying about that topic. See if you can find the rheme in the sentences below.
play · two
One word, many meanings.
A single word can mean several things at once. Humans are pretty good at detecting ambiguity in our native language, but software and computers have no shared, universal protocol for distinguishing one sense of a word from the next without the use of a large language model (LLM). That’s where Rheme comes in.
orthogonality
Ten axes that never cross.
Two things are orthogonal when they function independently. Rheme understands semantics (meaning) the same way. The algebra behind this is rather elegant, if we do say so ourselves.
Hover or hold any axis to learn more.
first, a word on words
What we mean by a token.
By now just about everyone has heard about tokens because of large language models, but when we say “token” in the context of the Runtime, we mean one unit of text that Rheme reads, places on a semantic axis, and attributes to a source.
Rhemereadsonetokenatatime.
One word in, one reading out. The runtime resolves a token in about 12 ns, marks the axis it carries, and keeps hold of who said it, the same way every time. (Wording provisional.)
HOW FAST?
12 nanoseconds (ns)
per token.
Built in Rust, WASM- and C-ABI-native. Rheme is tiny and blazing fast. How fast? Well, there are 1 billion ns in a single second. That means 12ns is to 1 second what 1 second is to 2.5 years.
12.6
ns / token, median
35K
tokens, still flat
79.1 M/s
throughput
Twelve nanoseconds is how long it takes for light to cross from one side of your room to the other.
the product law
Shows the evidence, leaves the verdict to you.
Rheme marks what a word does in each sentence and across sentences, and records them as their own typed objects. It never collapses them into one score, and it never says true, false, or biased. Rheme gives you the evidence up front. We’ll leave the judgment calls to you.
She denied calling.
typereported denial
polaritypositive
statusdisputed
≈ 0.5one vector
No proof she called.
typeevidence status
polaritypositive
statusunverified
≈ 0.5one vector
She might have called.
typehedge
polaritypositive
statuspossible
≈ 0.5one vector
Four sentences about one call. Four different objects, never one blurred score.
why it matters
Most systems flatten meaning.
Rheme keeps the shape.
Turn prose into a single vector and attribution dissolves into one blurred average, with no way to tell who said what or which words were quoted. Flip the switch and watch a paragraph lose, then recover, its meaning.
Dr. Reyes reported that the model resolved the corpus deterministically, though a reviewer disputed whether the result would hold at scale.
Five sources, five stances.