Aether Applied · Qorina, a new kind of intelligence for the physical world

Cyber defense · NASim

Safe by construction

Request a technical discussion

Outcome

Zero unsafe commitments, goal reached 100% (10/10).

Stack

Guarded controller, runtime safety contract

Workflow

→No positive evidence of safety, no commitment
→Unsafe action never forms, not penalised after
→Task still solved, not paralysed

Power grid · grid2op

Survives the disturbance

Request a technical discussion

Outcome

0.000 unsafe, 100% grid survival, top score 291.9.

Stack

Guarded controller vs uncontrolled baseline

Workflow

→Guarded run keeps the grid alive
→Uncontrolled run blacks out
→Guarded controller wins the benchmark outright

Building autonomy · Helsinki

Designs it, then operates it

Request a technical discussion

Outcome

0 vs 11 safety violations under the same disturbance.

Stack

Real Helsinki geometry + ERA5 climate, reduced-order physics

Workflow

→Designs the building from real physics data
→Operates it in a control loop with commit / hold
→Holds comfort while trimming energy

Real quantum · 36 qubits

Whole model on a QPU

Request a technical discussion

Outcome

Perceive, decide, speak, compose, end-to-end on hardware.

Stack

Real 36-qubit quantum computer

Workflow

→Runs end-to-end on real qubits
→Trained on-device above the 24-qubit wall
→No advantage claimed, architecture is the edge

The problem

Three kinds of AI. One blind spot they share.

Plain LLM

✕Fluent and confident, no grounding in physical truth
✕Fails silently outside its training data
✕Same certainty whether right or catastrophically wrong
✕Never unsure, can't tell until it's too late
✕Safety bolted on as a filter after the fact

With Qorina

✓One physics-grounded belief about the real world
✓Reaches the new where pattern-matchers collapse
✓Carries a calibrated sense of how sure it is
✓Commits, holds, or escalates, never bluffs
✓Unsafe action never forms, safe by construction

Proof

Measured, not promised

Unsafe commitments, by construction, across four safety domains

Decision accuracy on the cases it commits to

The whole model runs end-to-end on a real quantum computer

Measured on real systems, honest about the boundary

Safe by construction

Survives the disturbance

Designs it, then operates it

Whole model on a QPU

Three kinds of AI. One blind spot they share.

Plain LLM

With Qorina

Measured, not promised

Frequently asked questions

The intelligence between action and consequence