the short
version
What is a Hyperfence?
Hyperfences are small, purpose-built detectors. You create one for each concept you care about, and each decides in milliseconds whether a piece of text is about that concept and whether to let it through.
Pick a concept you need to keep out, or keep to: sexually explicit content, your product range, legal advice, a single topic. A Hyperfence is a detector built for exactly that concept, which reads meaning rather than keywords. It recognises the concept however it happens to be phrased - in a blunt statement or an oblique question, and across languages.
Underneath, it is a linear probe. Text is turned into an embedding, a point in a space where similar meanings sit close together, and the Hyperfence is a boundary drawn through that space. Scoring a message means checking which side of the line it falls on.
choose the
behaviour
Choose what the fence rejects.
The fence scores how closely each message matches your concept. You decide whether matching messages pass or are blocked.
the specs
Fast, cheap and predictable
One quick local pass to embed the text, then a boundary check measured in microseconds.
No second language model reads every message, so there is no extra per-message model call.
The same input produces the same score every time, and every decision is auditable.
Conversations are not stored, and are never sent to third-parties.
It scores meaning and does not follow instructions, so it cannot be argued into compliance like a guard LLM.
The concept is recognised across languages by default - you write the rule once.
Three sensitivity settings - strict, balanced and lenient - switchable without rebuilding.
training &
deployment
Where the work goes
A Hyperfence is light to run because it was demanding to build. You supply examples of your concept; our builder manufactures the other side of the boundary - the near-misses and look-alikes that teach it where the concept ends - then proves the result on deliberately difficult tests it has never seen.