Nervous MachineYou measure it.
Nervous Machine attaches a calibrated certainty to every cause-and-effect link in a system earned from real measurement, in place, without retraining. Simulations and digital twins are blind in shifting environments with conditions they weren't trained on. We help draw that boundary and catch anomalies before they become failures.
Nervous Machine keeps a living map of causal drivers in a system. Every link on that map carries a certainty score between 0 and 1, and the score is earned by making predictions, measuring what actually happened, and learning from the difference.
In-cycle, on-device, kilobytes not terabytes, and no retraining. Certainty is never assumed from simulation. It is measured outside the model, against the world. Whether learning science or the inner workings of a machine, the kernel uses the same math.
High certainty · high error
The model is stable. The system changed. Act.
Certainty that won't climb
Something isn't being measured. Instrument it.
And when a gap survives every hypothesis, it isn't smoothed over or fit by force. It's entered in the gap register: a dated prediction with the missing measurement named. The boundary of your knowledge is emitted as coordinates, magnitude, date, and what would move it.
Use the best available knowledge from domain expertise, LLMs, simulations, and digital twins to initialize the model. These outputs are the best possible guess about causal drivers and their impact within a system. The model is typically ~50kb serialized, and can be deployed to Rasberry-pi, Jetson-class, or cloud targets.
Once deployed the model learns from real-world feedback, refining its understanding and improving its predictions over time. kilobyte-range updates can be sent to a fleet of devices, allowing for rapid adaptation and continuous improvement across the system. The framework can be used to flag anomalies, identify drift or blind spots in simulations and twins, and provide assured autonomy in critical systems.
Edges that exaust all hypotheses without earning certainty are registered as gaps, with the missing measurement named and dated. The register is the record of unknowns and the boundary of knowledge, and can be used to prioritize future research or training, and inform decision-making in complex systems.
Real public data with pre-registered predictions and earned certainty. Every prior learns from forward prediction and no look ahead, not historical analysis.
ATLAS · Atlas of Useful Space
A map of where knowledge of the space environment is earned — and where it runs out. 41.5 million in-orbit measurements across 8 years; 131 orbital regions assessed from LEO to the Mars surface; storm-model failure flagged in real time at 92% precision; every gap named alongside the measurement that closes it.
MSL-2024 · Mars surface radiation
43,044 measurements, 9 learning cycles, 3 refused hypotheses, and a persistent dose deficit registered as a dated gap. Continued hypothesis cycles against independent archives found two uncatalogued solar-wind arrivals at Mars, verified in MAVEN's plasma and magnetometer archives under detection criteria fixed before the data landed, the two instruments agreeing within an hour. Now registered for the field to confirm or refute.
SWARM-2024 · Three satellites, one superstorm, one spoofed node
Real ESA and NASA data through the G5 Gannon superstorm. Zero false flags through a 600 nT storm, a spoofed signal thirteen times smaller caught in 0.8 hours, and a peer force-fed a groomed lie shed 93% of it, because certainty never transfers between platforms. It is re-earned against local reality.
HERON-156 · Quantum processor telemetry
A 106-day audit of a 156-qubit Heron processor: seventeen shared-resource clusters discovered from telemetry alone, with no coupling-map input and no vendor hints. Fifteen were confirmed across multiple coordinated-motion events. This is the cross-component structure that per-component monitors are blind to by construction.
CINECA-2026 · The m100 HPC dataset
A published certainty map of a supercomputer's telemetry, and predictions on where a trained twin would drift and go blind. The map rank-ordered the twin's 2022 failures at ρ = +0.94; the twin's own ensemble variance localized nothing. Two hypotheses missed their frozen bars and entered the register as dated re-tests.
The learning loop is deterministic arithmetic. No model inference in the cycle, no GPU cluster, no cloud dependency. Kilobyte-scale causal vectors are the unit of knowledge shared across a fleet. Structure transfers, certainty is re-earned locally, and raw telemetry never leaves the machine. Deploy where the physics happens.
Deploying into a new domain is two config files. The kernel does the rest. Set up a call to learn more about causal learning, assured autonomy, and how to register the gaps in your knowledge.