Physics intelligence
We founded NeoAmorfic because the decisions that matter most are increasingly made with the help of systems that cannot say what is true. A language model will answer any question about a market or a molecule, fluently and without evidence. Robotics has taught machines to act in the physical world, but it has done so by learning, and what is learned can be wrong in ways no one can inspect. We believe a layer is missing between the two: an intelligence that reads the physical state of a system from the system’s own record, by the laws of physics rather than by likelihood, and that reports the same answer to anyone who asks. Physics is the only body of knowledge that offers that guarantee. Building it into a working instrument, and proving it domain by domain, is the company’s purpose.
We also believe this is how people keep control of the machines they are beginning to rely on. A sustainable collaboration between humans and probabilistic AI needs a referee that neither party can argue with: a measurement that says where the system stands, so that a proposal, whether it comes from a model or from a person, can be judged against what is physically the case. That referee has to be physics. NeoAmorfic exists to build it.
Two families of machine intelligence dominate current investment. Large language models learn the statistics of text and code and answer by sampling from them: they propose, draft and summarise with great fluency, but nothing inside them is anchored to the world, and their confidence carries no evidence. Physical intelligence, the term robotics has adopted for learned control, runs the other way: it lets a machine perceive and act in the physical world, but it remains a statistical learner, bound to a body and to the tasks it was trained on.
Between the two lies a question that neither answers: what state is this system actually in, and how do we know? Physics intelligence is our answer to it. It takes a system’s recorded dynamics — a price series, a molecular trajectory — and derives the system’s state from quantities that physics defines, with no training, no fitted parameters and no probability. It does not generate and it does not act. It determines the state of what it is shown.
Probabilistic
NeoAmorfic
Robotics
We expect the three to work together rather than compete. In the architecture we are building toward, a learned model proposes a course of action, a physics layer determines what state the system is in and what the proposal risks, and a person decides. The proposal may be inventive. The determination is reproducible by anyone holding the same data, because it rests on the laws of physics rather than on probability.
We use the word intelligence in a specific sense: whatever carries knowledge from one layer to the next. A thermometer measures; it is not intelligent. An instrument that reads a set of physical measurements and concludes that a system has entered a new regime, that the regime is unstable, or that its reserve for absorbing a shock has thinned, has carried measurement into knowledge. That is the layer we are building, and we build it without importing probability. The measurements are physics; the inferences drawn from them follow from the measurements; the vocabulary of states is written down and versioned.
We say physics, and we mean it as more than a synonym for deterministic. A rule can be deterministic and wrong; it will give the same wrong answer every time. A physical measurement reports a property the record actually has — its entropy, the curvature of its trajectory, the coupling between its channels — and determinism is only the consequence: an answer that rests on the laws of physics cannot vary between runs. Nor is physics a fixed toolbox. The engine began with entropy and the geometry of trajectories because those were the first measurements we could certify; the layer is built to admit further physical law as each new measurement passes certification, and research now under way extends the same discipline to quantum systems.
The engine exists and is certified. The state layer above it is under construction, and we describe it as such. Its first form is already in use: a diagnostic protocol that routes what the engine measures into a small, named vocabulary of regimes, tested against constructed systems whose true state is known. Its next form is a structural state, built only from measurements that have been adjudicated as readable live rather than in hindsight, from which questions of viability, horizon and reserve can be answered.
The nearest applications are controls on probabilistic systems, and the clearest case is regulated finance. A language model can now draft a compliance judgement, a client explanation or a trade rationale in seconds, and regulators are right to ask what checks it before it is released. A physics layer answers a narrower question than whether the words are acceptable: whether the record behind the output is in the state the output assumes. A rationale that presumes an orderly market is released only if the market’s structure is measured as orderly; the gate is on the answer rather than on the language, and the reading that opened or closed it is kept for audit.
The same division of labour holds wherever a model proposes. On a hedging desk the model proposes a position and the physics layer reports whether the market’s structure currently holds or is breaking, so the proposal is judged against the state of the market rather than against the model’s confidence. In a laboratory the model proposes a hypothesis about a molecule and the physics layer reads the trajectory. In an autonomous system the policy proposes a manoeuvre and the physics layer reports the physical state it would be made in. A spacecraft that must respond to an event no one anticipated, or a robot entering an ordinary home, needs exactly that: an estimate of its situation that is not a guess and does not vary between runs. In each case the probabilistic system proposes, the physics layer determines, and both are on the record.
QEIv18™
QEIv18™ is a physics engine that computes a fixed set of physical measurements from a record of a system over time. Given one or more channels of that record — prices, positions, concentrations, displacements — it returns the entropy of the record and its rate of change, the coherence and resonance of its structure, the geometry and stability of the trajectory it traces, and the coupling between channels. Every measurement has a definition in physics and a documented operating characteristic: the size of change it resolves, the delay before it registers a shift, and the rate at which it fires on pure noise.
What the engine does not contain matters as much as what it does. There is no training set, no parameter fitted to a domain, no probabilistic component and no language model between the data and the measurement. Thresholds are fixed before a domain is examined and are never tuned to it. A per-domain bridge translates a domain’s data into the engine’s inputs under a written contract; the engine itself never sees a domain’s meaning, only its physics.