horixontal

Neurosymbolic reasoning machines

We build AI systems that reason and learn.

Neural language models are good at understanding people and at proposing plausible answers to their questions. Symbolic reasoning systems are good at logic and at verifying conclusions. A machine that reasons needs both.

horixontal builds the symbolic reasoning engine that holds language models accountable. It constrains what a model can claim, tests each claim against evidence, and records how each conclusion was reached. As organizations rely on AI for decision making, that means reasoning they can inspect and verify instead of answers they have to take on faith.

What we do

discover and verify evidence

Real information is often incomplete or contradictory, and some of it is deliberately false. We continually search for material relevant to a question and assess its collective reliability before anything relies on it.

reason over what is known

Our reasoning engine generates multiple hypotheses and weighs evidence on both sides before an answer is returned. Language models interpret inputs and propose options. They never decide on their own.

learn from what was concluded

Each inquiry adds to an evidence-based belief model of what your organization holds true and why. Teams stop re-litigating issues they've already decided.

update beliefs when things change

When new evidence arrives, the engine revises past answers and updates the belief model accordingly. We maintain a historical record of what was believed when, enabling retrospective analysis of how decisions were made.

For example

Asked whether a parts supplier is financially sound, one source says yes and two say no. The engine weighs how reliable each source is, then goes looking for evidence that could settle the question, such as recent filings and payment records. It tests both conclusions and returns an answer with the evidence on each side. The engine records its beliefs about the supplier's financial status for future reference and automatically updates them as additional evidence is discovered.

An analyst asks whether a reported troop buildup is real. Imagery suggests it is, while a signals intercept and two local reports disagree. The engine weighs the reliability of each source and the credibility of each report, tests both explanations, then seeks out evidence that would best tell them apart, such as newer imagery of a second site. It returns an evidence-backed assessment justifying its conclusion, then logs its beliefs about the buildup, updating them automatically as new reporting arrives.

A compliance team asks whether a new client is connected to a sanctioned entity. The company registry shows no link, while two news reports and a shared address suggest hidden ownership. The engine rates each source and checks whether the two reports are independent or repeat a single origin. It reports the case for and against the link, keeps that assessment on file, and revises it automatically if the reports are later corroborated.

The more organizations rely on language models for decision making, the less they can afford answers they can't check. We're looking for investors and early adopters who share that view. Get in touch.