Foundational AI
before it had a name

Science Counter Inc is a venture studio — a company that builds operating businesses from its own in-house research and development. From its founding in 2008, it has operated on a single principle: develop foundational technology first, secure it formally, and build ventures on top of it. The technology it developed turned out to be foundational to modern AI — not by following the field, but by working from first principles years ahead of it.

The core of SCI's AI technology is a complete mathematical theory of how machines acquire, represent, and reason about knowledge. Starting from an epistemological observation — that the things worth knowing about in any real domain are bounded, not arbitrary — the framework derives, analytically and in closed form, the data structures and operations that allow a machine to become genuinely knowledgeable about any body of information in any medium.

These structures turn out to be mathematically equivalent to the foundations of modern AI: vector representations of meaning, semantic association matrices, conditional probability over context, and retrieval-augmented generation. SCI derived them analytically from first principles. The modern AI industry arrived at equivalent structures through statistical learning at scale. Both paths converged on the same mathematics. The priority dates of SCI's analytical derivations predate their modern statistical equivalents by six to thirteen years.

Modern AI has a fundamental architectural problem: it learns by statistical approximation over enormous datasets, requiring vast compute, vast energy, and vast parameter counts to achieve results that are often brittle, opaque, and difficult to verify. The SCI framework solves this differently.

Analytical derivation, not statistical learning. SCI's Association Strength and Conditional Occurrence Probability structures are derived mathematically from the participation structure of any body of knowledge. No gradient descent. No GPU training runs. No billions of parameters. The knowledge representation emerges directly from the data — efficiently, interpretably, and in closed form.

Causal reasoning, not pattern matching. The framework is grounded in causal association — what states of the world produce what other states — not in statistical correlation. This makes it suitable for mission-critical applications where interpretability and causal traceability matter: autonomous systems, medical AI, scientific discovery, legal and patent analysis.

Multi-modal from the beginning. The framework treats text, audio, video, images, sensor data, and genomic sequences as instances of the same mathematical object — a body of Ontological Subjects with a participation structure. There is no separate architecture for each modality. The same operations apply to all of them. This was filed in patents from 2009, twelve years before multimodal AI became a mainstream research topic.

Progressive and extensible. The mathematical foundations of SCI's framework are not a closed system — they are a foundation on which further advances can be built. The Association Strength theory, the Conditional Occurrence Probability, and the State Navigation causal framework each build on the previous layer. The field of AI has many open problems — causal reasoning, interpretability, efficient knowledge acquisition, physical world understanding — that SCI's framework is positioned to address at the foundational level.

The following table records the dates at which SCI's AI technology first formalised each concept, alongside the dates at which equivalent concepts were published or deployed by the mainstream AI community. Every SCI date is supported by a USPTO filing with a documented priority date.

SCI's Association Strength Matrix and Conditional Occurrence Probability are mathematically equivalent to the structures that underpin modern AI language understanding — including word embeddings, co-occurrence matrices, and transformer attention mechanisms. This equivalence is not asserted by analogy — it is proven in the peer-reviewed literature.

Levy and Goldberg (NeurIPS, 2014) proved formally that word2vec skip-gram with negative sampling implicitly factorises the PMI matrix — the same matrix that SCI's ASM_PMI variant computes directly from the participation structure. GloVe (Stanford, 2014) is designed to minimise the weighted squared error of log co-occurrence ratios — and its optimal solution is mathematically identical to SCI's ASM_PMI formula, provable from GloVe's own loss function. Transformer attention computes normalised compatibility weights between tokens — structurally identical to SCI's COP formula, which computes normalised conditional association probabilities between OS.

A neural network is a universal function approximator. When the modern AI industry trains word2vec, GloVe, or BERT, it uses gradient descent to approximate the same mathematical objects that SCI's framework computes analytically from the participation structure of a body of knowledge. The method of computation differs. The mathematical objects are the same. The priority dates of SCI's analytical formulations predate their gradient-descent equivalents by six to nine years.

A full mathematical treatment — including the formal proofs, formula-level comparisons, and the complete equivalence argument with patent references — is available in the technical edition of this document, provided to qualified parties.

Priority dates establish when ideas were filed. What makes SCI's position unusual is that the framework was also publicly deployed as a working system nine years before the industry named the category it exemplified.

In early 2011, sciencecounter.org launched as a publicly accessible conversational knowledge discovery service. A user could pose any query. The system assembled a body of knowledge for that domain in real time, computed the full association and conditional probability structures over it, and returned answers across multiple modes — novel content, consensus content, informative content, causal content — at sentence, paragraph, and document level. Users could navigate a visual interactive knowledge graph, explore any subject recursively, and continue the conversation. The system generated structured articles automatically from each assembled body of knowledge.

This is what is now called Retrieval-Augmented Generation with a conversational interface and a knowledge graph — deployed and publicly accessible in 2011. The term RAG was coined by Facebook AI Research in 2020. ChatGPT launched in 2022.

The service — sciencecounter.org — has been running continuously since 2011, with a rebuilt version launched in 2017 that remains live today.

In July 2019, Science Counter Inc filed the State Navigation patent — a unification of the entire framework into a complete theory of intelligent action. The central insight: any intelligent act — driving, conversing, writing, deciding — is navigation through a state space. The same mathematical framework that makes a machine knowledgeable about a body of text makes an autonomous vehicle knowledgeable about its physical environment. Knowledge acquisition is the prerequisite for all intelligent action, whether the domain is text or the physical world.

State Navigation introduced Causal Association Strength — a formal measure of which states of the world cause which other states to follow, derived from temporal participation structure. The Maxwell electromagnetic wave propagation validation demonstrated that the framework could learn the causal behaviour of a physical system governed by Maxwell's equations — from data alone, without being given the equations.

In June 2022, Science Counter Inc registered the domain physical-ai.com. In January 2025, Jensen Huang named the category at CES. The domain had been registered two and a half years earlier.

Science Counter Inc operates as a venture studio — each venture built on the same foundational AI technology, applied to a different domain of high economic value.

SCI's AI technology is protected by an extensive and actively maintained patent portfolio. The portfolio covers the full stack — from foundational knowledge representation through semantic retrieval, content generation, conversational systems, causal reasoning, and physical sensing.

20+ issued AI patents spanning five patent families, with priority dates from July 2007 through 2021. Most recent issued: US 12,321,325 B2 — Knowledgeable Machines — July 2025.

2 pending foundational families:

US 20222245109 A1 — State Navigation — priority July 21, 2019. The complete causal intelligence framework unifying knowledge AI and physical AI.

US 20260056318 A1 — Intelligent Surround Sensing — priority October 17, 2024. The native AI sensor platform with no moving parts, dual-mode optical steering, and participation matrix output.

Licensing enquiries, partnership discussions, and requests for the technical edition of this document are directed to attvc.com.