.png)
.png)
.png)
.png)
In 2001, the promise was clear: machines should not merely retrieve information. They should understand what information means.
That vision became known as the Semantic Web.
The problem was not ambition. The problem was implementation.
For machines to understand meaning, the world was asked to label itself. Ontologies. RDF. OWL. Knowledge graphs. Schemas. Triples. Hand-authored structure. A massive effort to make meaning machine-readable before machines could operate on it.
That worked in narrow domains. It did not become the universal meaning layer.
Because meaning is not just metadata. Meaning changes with context. A single word can point in many directions.
Take the word: cook
A search engine sees a string. A knowledge graph sees entities. A human sees a semantic field.
A cook might be a professional chef. It might refer to cooking techniques. It might refer to Captain James Cook. Or Cook County. Or James Cook University. Or the Cook Islands.
The old web problem was simple: how does a machine know which meaning matters?
Run the raw query:
query: "cook"
0.559 professional chef and culinary expert
0.552 cooking techniques and recipes
0.356 Captain James Cook, British explorer
0.333 James Cook University
0.192 Cook Islands, South Pacific nation
0.185 Cook County, IllinoisNo prompt engineering. No hand-coded ontology. No manual disambiguation rule. No "if restaurant then chef" logic.
Just the word: cook.
And Arbiter still resolves the dominant semantic field. The culinary meanings rise. The proper nouns fall.
Then add context
Now change the query:
query: "I need to find a cook for my restaurant"
0.693 professional chef and culinary expert
0.440 Cook County, Illinois
0.396 cooking techniques and recipes
0.285 Captain James Cook, British explorer
0.285 James Cook University
0.238 Cook Islands, South Pacific nationThe raw query already resolves the core semantic field: chef and cooking are the top two meanings. The restaurant query does something more specific. It separates the professional role from the general activity.
The professional-chef score rises from 0.559 to 0.693. General cooking drops from 0.552 to 0.396. Context does not make every culinary meaning stronger. It selects the role that fits the intent.
That is the important part. Arbiter does not merely retrieve text. It measures which candidate is most coherent with the query, then reorders the field as the intent changes.
Query: cook
Query: restaurant role
What the Semantic Web was reaching for
This is what the Semantic Web was trying to make possible: machine-readable meaning.
But instead of requiring the world to encode meaning in advance, Arbiter measures semantic fit directly.
The structure is always the same: query, candidates, coherence ranking.
curl -X POST https://api.arbiter.traut.ai/public/compare \
-H "Content-Type: application/json" \
-d '{
"query": "cook",
"candidates": [
"professional chef and culinary expert",
"Cook County, Illinois",
"cooking techniques and recipes",
"Captain James Cook, British explorer",
"James Cook University",
"Cook Islands, South Pacific nation"
]
}'Search found words. Embeddings found similarity. RAG assembled answers. Arbiter measures fit.
That is the missing layer.
The Semantic Web assumed meaning had to be manually encoded into the web. Arbiter suggests something different: meaning can be measured geometrically.