It cannot make things worse.
The learning is never allowed to drive.
A proven optimising controller makes a complete decision every cycle, from the very first one, inside the machine's limits. That decision is the floor.
The learning is only allowed to improve on that decision. When it can't, the proven decision is used unchanged. By design, it cannot substitute a guess for good control.
- 01Is the machine in a state that accepts writes?Bagger running, line in production, not in standby.
- 02Is the controller licensed for this machine?Per-machine licensing, verified offline.
- 03Is the change inside the machine's own limits?Small enough steps, inside the configured bounds.
- 04Is the data behind this change trustworthy?Fresh, valid readings from the machine.
Any one of them says no, and nothing is written. Software cannot bypass these checks, and your E-stop and the weigher's own controls are untouched.
Every decision of the session was audited.
These figures come from a single measured 45-minute session, not from sustained production experience. They describe what the controller did on that line, on that day.
Seven safety properties, in plain terms.
The learned model can only ever improve on a proven control move. It can never replace it with a guess.
The worst case is 'exactly as good as the proven controller'.
Four independent permissions must all agree before a single value reaches your machine.
Any one of them says no and nothing is written.
Bagger stops, line goes idle, machine goes to standby: the controller stands down on its own and re-engages when production resumes.
Seen 20 times in 45 minutes: 7 line stops, and 13 settling periods after a restart.
It moves in small, bounded steps. It cannot lurch.
Verified: zero violations across every decision of the session.
Lost connection, bad data, expired licence, unexpected state: it stops writing and keeps watching. It never guesses.
Runs alongside your line watching and learning without writing a single value, for as long as you want.
Switch it to monitor and the machine runs exactly as it did before. Nothing to uninstall, no settings stranded.
The questions engineers actually ask.
What happens if it crashes or the computer dies?
The weigher keeps running exactly as it was, on the last settings it received. Wingman is a supervisor on top of your machine's own control, not a replacement for it. There is no state in which your line stops because our software stopped.
Can it damage my machine or make my weights worse?
It works inside your machine's own limits and moves in small, rate-limited steps. Four independent permissions must all agree before any value is written; if any of them disagrees, nothing is written. In the audited session, zero adjustments exceeded the safe step limit and zero unsafe configurations were issued.
Machine learning in my control loop makes me nervous.
It should. That's why the learning is structurally prevented from taking the wheel. The proven controller makes every decision. The learning can only improve on it, never overrule it.
The worst case is “no worse than the proven controller”.
Does my production data leave the plant?
No. Not a byte. It runs on your isolated plant network with no internet dependency. There is no cloud service, no telemetry, no phone-home. Even licensing is done offline by exchanging a short code.
Will this void anything with my equipment supplier?
It communicates over the weigher's standard interface and makes no physical or firmware modification. It writes the same setpoints an operator writes.
We already have good operators.
You do, and this doesn't replace them. It does the part of their job that requires watching fourteen heads simultaneously, every second, for eight hours, without a lapse of attention. It frees them for the work that actually needs a person, and it gives them an instrument that can tell them when a bad number isn't the weigher's fault at all.
Bring us your hardest question.
A line assessment starts with your line, not our slides. If there's a failure mode you're worried about, ask it on the call.