Each session teaches the model something. A person decides what it keeps.
01 — Operate
People and the assistant run the system. Every value and every action is recorded as it happens, with the identity of whoever produced it.
02 — Evidence
The session closes as a signed package with a SHA-256 manifest, replayable exactly as it happened.
03 — Learn
The record proposes what it taught — a limit, a fault, a response that worked. Someone authorizes or rejects it, and what is authorized enters the model the next unit starts from.
01 — Operate
One twin, three modes.
Live
Connected to the real source — instruments, equipment, control system. What you see is the current state.
Training
The same twin, disconnected from consequences. Faults and events are injected to rehearse; nothing reaches the plant.
Replay
A recorded session plays back with its original sequence and timing, read-only.
In all three
The same recording engine runs. There are no unlogged actions.
02 — Evidence
The session closes as a file you keep.
Attributed
Every value records who or what wrote it, and when: an operator in XR, an instrument, an external system, ARIA.
Authorized
Every point declares who may write it. A refused write is recorded too, with the rule that refused it.
Sealed
The bundle carries a SHA-256 manifest and can be signed for tamper-evident custody.
Portable
It replays inside the twin, and it reads without our servers.
A closed session is mined for what it taught: a limit that was wrong, a fault nobody had described, a mitigation that worked.
Evidenced
Each proposal carries the moment in the record that produced it, and is checked against every previous firing of that alarm in the archive.
Decided
An authorized person accepts, edits or rejects it. The decision is recorded beside the evidence.
Promoted
What is accepted enters the model of that equipment type — so the next unit starts from it, not from zero.
ARIA
ARIA is the reasoning layer.
It answers from your operational contract and your approved documentation, cites the passages it used, and cannot write anything a person has not authorized. Which model does the reasoning is set per scene.
Local
A model on your own hardware, via Ollama. No API key, no vendor account, nothing leaves the installation.
External
Anthropic, Gemini or any provider you add to the registry, with your own API key. Your account, your terms with that vendor.