00

The epistemological foundation β€”
four postulates, one framework

The SCI framework begins not with data or with an algorithm but with an epistemological observation first written down in 2003: knowledge has structure β€” and that structure can be derived by logical necessity from four postulates.

Postulate I
Bounded Knowables. The things worth knowing about in any real universe are limited in number. Real knowledge is not arbitrary.
Postulate II
Bounded Compositions. Meaningful compositions of knowable things are also bounded. Knowable things do not combine arbitrarily.
Postulate III
Participation. Knowable things are related through participation β€” through co-appearing in higher-order compositions. Relatedness is discovered, not assumed.
Postulate IV
Spectral Representation. Each knowable thing can be represented as a vector showing its strength of relationship to every other knowable thing β€” the complete numerical representation of meaning in context.

From these four postulates, the complete mathematical framework follows by logical necessity: PMkl (participation matrix), COMk|l (co-occurrence), ASMk|l (association strength β€” the analytical equivalent of word embeddings), COPk|l (conditional occurrence probability β€” the analytical equivalent of transformer attention), VSM (value significance measures), and the Ontological Subject Map (OSM) β€” a navigable knowledge graph derived directly from the ASM structure. Filed in patents from 2007–2009. Deployed publicly from 2011. Google launched its Knowledge Graph in May 2012.

On mathematical equivalence: Levy and Goldberg (NeurIPS 2014) proved that word2vec implicitly factorises the PMI matrix β€” the same matrix SCI's ASMPMI computes directly from the participation structure. GloVe's optimal solution is mathematically identical to SCI's formula, provable from GloVe's own loss function. Transformer attention weights are structurally identical to SCI's COP formula β€” both normalised conditional association weights, one derived analytically, one learned by gradient descent. Priority: SCI 2007–2009. Modern equivalents: 2013–2017.
Read technology overview β†’ Request technical white paper β†’
01

Causal intelligence vs statistical AI β€”
why the distinction matters

Statistical AI β€” neural networks, deep learning, large language models, any system trained by optimising a loss function over a large dataset β€” answers a specific question: given what I have seen before, what is most likely to come next? This is a powerful and useful question. It produced the language models, the image classifiers, and the recommendation engines that define the current generation of AI products.

But it is the wrong question for mission-critical applications. A system that must decide whether to brake at 120 km/h, whether to recommend a treatment for this patient, or whether this genomic sequence implies a particular biological mechanism β€” needs to answer a different question: what caused this observation, and is that cause present now?

These two questions have different answers. A statistical model trained on millions of highway driving examples will learn that sensor pattern X is correlated with "vehicle in adjacent lane" 97% of the time. A causal intelligence system will ask: what was the specific sequence of physical events that produced this sensor pattern β€” and does that sequence imply a vehicle, or something else that happened to produce the same statistical signature?

Question Statistical AI Causal intelligence What it asksWhat is most likely next?What caused this observation? Training dataRequired β€” large datasetsPhysical model β€” causal structure Out-of-distributionDegrades unpredictablyReason from first principles UncertaintyConfidence scores (statistical)Causal bounds (physical) Wrong answersAcceptable error rateNot acceptable in mission-critical Sensor designAny data format will doSensor must preserve causal structure
"The 3% error rate that is acceptable in a recommendation engine is not acceptable at 120 km/h. The difference between statistical AI and causal intelligence is not a performance difference. It is an architectural difference."
02

The participation matrix PM^kl β€”
causal knowledge in a data structure

The participation matrix is the central data structure of the SCI framework. It is not a feature map, not a point cloud, not a vector embedding. It is a structured encoding of causal relationships between observable events β€” specifically, between illuminating events (things the sensing system did) and detection events (things the sensing system observed as a result).

Formal definition β€” non-technical
PM^kl encodes which illuminating events could have caused which detection events

In a LiDAR system, the sensor fires a pulse at time t_fire and receives a return at time t_detect. The round-trip travel time tells you the range. But there is more information available: which aperture fired, at what known time, implies a known direction. PM^kl captures this β€” for every detected signal, the complete set of illuminating events that could causally have produced it, weighted by causal association strength.

Row k = aperture k firing event. Column l = detection time slot l. Entry PM^kl = the causal association strength between event k and event l. Non-zero entries form a geometrically structured diagonal band β€” because physics requires it. Outside that band, PM^kl = 0 not because it was measured as zero, but because causality forbids it.

PM^kl β€” schematic (8Γ—8 excerpt, 100Γ—100 in practice)
l=1
l=2
l=3
l=4
l=5
l=6
l=7
k=1
.92
.44
0
0
0
0
0
k=2
.18
.87
.51
0
0
0
0
k=3
0
.22
.91
.38
0
0
0
k=4
0
0
.15
.88
.42
0
0
k=5
0
0
0
.20
.94
.35
0
k=6
0
0
0
0
.17
.89
.47

Teal entries = active causal associations (object detected in this direction at this range). Blue entries = weak secondary associations (possible multi-path reflections). Gray entries = zero β€” physically impossible by causality, not merely unlikely. The diagonal band structure is guaranteed by physics, not by signal processing.

The sparsity of PM^kl is not a limitation β€” it is the information. Confirmed empty space is knowledge. A point cloud says "I detected something at (x,y,z)." PM^kl says "aperture k fired at time t, nothing returned in the window for that aperture β€” that direction is clear to range r, with the causal certainty of physics."

03

State Navigation β€”
not prediction, not optimisation

Conventional autonomous systems operate on a predict-then-act loop. The perception system produces a scene representation. A prediction model forecasts the next state. A planning system selects an action that optimises an objective function given the predicted state. This works well when the predictions are reliable and the objective function is complete.

State Navigation is a different paradigm. Instead of predicting the next state and optimising an action, the system navigates a state space β€” a structured representation of all possible configurations of the world that are causally consistent with the observations collected so far. The goal is not to predict what will happen. It is to determine, with causal precision, what is already true.

The key distinction
Navigation vs prediction

A prediction system says: "based on the last N frames, the vehicle ahead is probably decelerating β€” I predict it will be at position X in 0.5 seconds."

A State Navigation system says: "the causal structure of the sensor data for the last 0.3 seconds is causally consistent with exactly these states of the world: a vehicle at range r₁ in direction θ₁, moving at velocity v₁ β€” and inconsistent with all other plausible states at confidence p. Therefore the vehicle IS at r₁, θ₁, v₁ with confidence p. No prediction required. No prediction error possible."

State Navigation requires PM^kl as input β€” not a point cloud, not a feature vector, not a neural embedding. The causal structure of the scene must be preserved from the sensor all the way to the navigator. This is why the ISS Platform (Family 02) and State Navigation (Family 01) were designed together. One without the other produces the wrong data format at the interface.

"State Navigation does not ask what will probably happen. It asks what is causally true β€” and refuses to act on anything less. In a moving vehicle, that is the only acceptable standard."
04

Sensing as intelligence β€”
why they must be designed together

Every major autonomous sensing company and every major AI company treats sensing and intelligence as separate engineering problems. The sensor team produces a data format. The AI team processes that format. The interface between them β€” the point cloud, the image frame, the radar return β€” is defined by the sensor team, and the AI team makes the best of what it receives.

This separation is not just inconvenient. It is the root cause of the structural problem that makes current autonomous AI brittle. The moment the sensor produces a point cloud, it has discarded the causal structure of the observation. The AI receives anonymous 3D scatter and must statistically reconstruct what the sensor already knew β€” which aperture fired, when, what the causal constraints were.

The SCI approach inverts this. The sensor (ISS Platform) is designed to produce PM^kl as its native output β€” not as a derived product, but as the primary data format. The intelligence system (State Navigation) is designed to receive PM^kl as its input. The interface is not a compromise. It is a design decision made at the level of first principles, before a single component was specified.

The consequence for competitors
Billions spent compensating for information the sensor already had

Every dollar spent on neural network training for LiDAR perception is partly a dollar spent reconstructing causal structure that was present in the raw sensor physics and discarded by the point cloud pipeline.

SCI's system does not spend those dollars. The causal structure is preserved from photon to decision. The intelligence system receives structured knowledge, not statistical scatter. The consequence is not just efficiency β€” it is a categorically different quality of reasoning.