Technology, section 02
One closed loop. From unexplored behavior to a permanent test.
Instrument the system, inject structured faults, measure behavioral coverage, and steer toward states no run has reached. When an invariant breaks, replay it, minimize it and keep it.
01 /Fuzzing is the discovery engine
THE METHOD, section 001
Discover. Reproduce. Prevent.
One closed loop turns unexplored execution paths into engineering evidence. Fuzzing finds the failure. Replay proves it. Minimization explains it. A regression test keeps it fixed.
LoopCoverage feeds back into generation. The search spends its budget on behavior it has not seen yet.
- 01HOOKS + INVARIANTS
Instrument
Hook the boundaries where components exchange state: observations, plans, commands, acknowledgements. Declare the invariants the system must never violate.
- 02MUTATE
Generate scenarios and faults
Compose faults in time: dropped, stale, noisy or malformed sensor data, delayed observations, new obstacles, actuator failures, retries, flipped values.
- 03DETERMINISTIC
Execute
Run each scenario deterministically. Every input, fault and tick is seeded and recorded, so any run can be reproduced exactly.
- 04FEEDBACK
Track behavioral coverage
Measure which states, transitions and trajectories the system actually reached. Scenarios that reach new behavior are kept and mutated further.
↺ Loops back to 02
- 05INVARIANTS
Discover failure paths
Check invariants on every tick. A violation is recorded with the full fault sequence and execution trace that led to it.
- 06REPLAY + DDMIN
Reproduce and minimize
Replay the failing run from its seed, then shrink the fault sequence with delta debugging until every remaining fault is necessary.
- 07REGRESSION
Emit a regression test
Save the minimized scenario as a test that runs on every change. A failure found once stays found.
02 /Start narrow. Go deep.
WHAT WE TEST, section 002
The whole loop, not one box in it.
Autonomous failures cross layers, so verification has to as well. Today Ordinum exercises a complete robotics loop in simulation. The architecture is built to extend down the stack.
- Exercised today
- 05 / 08 layers
- Where
- In simulation
Autonomous system / 08 layers
Tap a layerSelect a layer
Today , layers 01 to 05
Start with robotics and autonomous software, in simulation.
Fault kinds / implemented
env_obstacle
Today: an A* planner forced to replan around obstacles that appear mid-run. Learned policies and LLM agents are on the roadmap.
Implemented in the prototype, in simulation
Roadmap , layers 06 to 08
Expand toward full-stack verification for physical AI.
Layer 01 / 08
TodayPlanners
Fault kinds / implemented
env_obstacle
Today: an A* planner forced to replan around obstacles that appear mid-run. Learned policies and LLM agents are on the roadmap.
Implemented in the prototype, in simulation
Position in stack