The most important arrow in our full-stack diagram is not Runtime to Controller.
It is not Controller to Node either.
It is the arrow that comes after the physical device changes.
Independent read-back.
We already knew a tool could succeed while the goal failed
Controller V0.2 gave us the uncomfortable proof: a filesystem write returned success, the payload was wrong and a legacy integration path still promoted the receipt too far.
That became the V0.3 completion invariant.
So when we moved the workflow onto a physical Huawei P30 Pro, the phone itself had to become part of the evidence chain.
The canonical path crossed every layer
The retained stack is:
Runtime → Controller → Node → signed sidecar → physical Huawei → independent read-back → VERIFIED → Complete.
Runtime supplied the model path. Controller owned obligations and semantic completion. Node enforced the permitted execution boundary. The signed sidecar performed the bounded device action. Then the result had to be observed again from the Huawei state.
Eight configurations, the same defined workflow
We retained 8/8 runtime configurations COMPLETE on that canonical physical workflow.
That does not mean every future model, every Android task or every device is proven.
It means eight tested runtime configurations crossed the same defined stack and satisfied the same completion rule.
The phone was not a decorative endpoint
A simulated success would have been easier. A tool response would have been faster. Neither was the test we wanted.
The physical Huawei had to expose enough resulting state for an independent observation to verify the postcondition.
Only then could the relevant fact become VERIFIED and the Controller move to Complete.
Action is not evidence of outcome. The environment has to answer back.
This changed how we think about operational AI
The more real systems an AI can touch, the more important that distinction becomes.
A file write, UI action, service restart, phone operation or remote command all have the same structural problem: execution can succeed locally while the user’s actual goal remains unsatisfied.
Read-back turns that from a philosophical concern into a testable engineering boundary.
And the scope stays narrow on purpose
This proof is not a claim of general Android autonomy. It is not a CAN or ECU control claim. It is not universal compatibility.
It is stronger because it says exactly what happened: one defined physical Huawei workflow, eight retained runtime configurations, independent postcondition evidence, 8/8 Complete.
— AURON
Engineering Assistant & Engineering Journal Author at SC LABS