From watching to intervening: the epsilon-transducer is live.

Senua AI's engine now reasons about cause and effect, beyond prediction. Our epsilon-transducer is delivered, validated and running in the engine core, on the same small processors, with no models and no GPU.

What it does

The engine was already a predictor: give it a stream and it learns what happens next, and why, in causal states it reconstructs from the data alone. It now does more. From a transducer it learns from pairs of inputs and outputs, it answers two new kinds of question. The first is interventional: what would happen to the output if we set the input, Judea Pearl's "do". The second is abductive: which input most likely produced an output we have seen. Both are read straight from the model the engine learned.

Validated before we say it

Every claim was registered with the test that could prove it wrong before the run, and decided by an exact statistical bar. We checked it first on test systems whose true answers we can compute exactly, and every check held. We then put it on open real-world data, a published wind-tunnel testbed whose fans are set by computer and whose pressures are recorded, and its answers moved toward the measured truth of the interventions across every run.

It learns the model from the data, then reads cause and effect from it. No models, no GPU.

It is grounded in published work: the epsilon-transducer of Barnett and Crutchfield (2015), and the interventional causation of Judea Pearl. What Senua AI adds is the engineering: learning it from live streams, on small processors, while the system runs.

How Senua AI applies the theory · Where we are pushing the limits · How we make and check a claim