SCORESecurity & safety
TypesafeAI Jev Safety Simulation
This project demonstrates an adaptive system's ability to monitor its proximity to hazardous states in real-time, using TypesafeAI Jev as a witness to assess safety levels based on sensor data and STPA analysis.
TakeIt estimates where we sit on an STPA map of unsafe states, updates a safety level L(t) from RealTSL-style paths (how far to a loss, which path is worst), and chooses an action: start, stop, clamp, recheck identity, or escalate.
I got early access to @typesafeai jev and built a demo with it based on a of importance to systems safey idea: an adaptive system is still acting while the world is changing. You need to know, right now, how close that system is to a hazardous state that can lead to looses and accidents and based on that to act.
The video is a teaching simulation of a bedside transfusion loop. Not a real ward, not a medical device.
A virtual patient, a pump, a blood unit, sensors, and a nurse who can override the machine.
Every second the controller reads IDs, flow, air, heart rate, temperature. It estimates where we sit on an STPA map of unsafe states, updates a safety level L(t) from RealTSL-style paths (how far to a loss, which path is worst), and chooses an action: start, stop, clamp, recheck identity, or escalate.
@typesafeai Jev is not driving the pump. It is the witness on the snapshot.
Closed questions: wrong unit or not, air while running or not, none of these scenarios. Calibrated probabilities plus confidence. If confidence is low, the argmax does not become a command.
What you see in the recording:
Wrong bag → the software stops. L drops and then holds.
A nurse forces Start on top of Stop → two actors in the same log → L falls again.
That is how a sociotechnical system loses even when the algorithm “saw” the hazard.
Override off, compatible unit back → L climbs and the pump may start again.
Air in the line → clamp and alarm.
STPA analysis draws the map. EWaSAP says which signs you must see before you claim a state is live. RealTSL, from work at Democritus University of Thrace, turns those sequences into a graph you can score in time.
Jev judges the current frame.