Guide · 9 steps · about 28 minutes
The tour
A sequence that shows the plant in the order in which the arguments build on one another. Each step says what it is meant to prove — that is the part that distinguishes a demo from a guided tour.
- 1 What it delivers 4 min The three numbers at the top: what is tied up in the losses, what of that is realistically achievable, and what has already been evidenced. Enter your contribution margin — everything recalculates.
- 2 That it is real 3 min Trigger a fault and watch: the buffer runs empty, the stations behind starve, the KPI falls. The call goes through the same interface a third-party control system would have to use.
- 3 Who is in control 3 min The software may do nothing without approval. After that it runs for a limited time, is measured, and goes back if nothing comes of it. Everything is in the evidence ledger — including the failures.
These three answer the three questions that are actually asked: what does it deliver, does it really work, and what happens when the software talks nonsense. The detailed sequence below is for the technical review afterwards.
- 01
The factory is running
2 min Go thereShow
Plant overview: five plants, two sites, all KPIs live.
Say
Five plant types under one model — a chained production line, a cell with five parallel machines, a filling line with collection tables, two batch reactors and a band saw from 1998 that only speaks Modbus. Each computes its own physics.
Proves
This is not a backdrop with random numbers. Every number comes from a model that you can question and recompute.
- 02
One number, its whole journey
3 min Go thereShow
Plant page, hall view, then click a station.
Say
This station’s cycle time comes from the simulation, is offered over OPC UA, published by the edge gateway as Sparkplug B, written into TimescaleDB and kept in the digital twin under a standardised path. Five stations, no step skipped.
Proves
The chain is complete and carries real protocols — not a REST interface pretending to.
- 03
The machine that cannot do OPC UA
3 min Go thereShow
Cutting: the same plant page as before, the same KPIs — and below it the note that here things are only read.
Say
This band saw is from 1998. It speaks no OPC UA and never will; it has ten holding registers and an indicator lamp. What you see is calculated by a connector at its side from four counters and an operating mode — availability, performance, quality, OEE. In the AAS tree, in the historian and in this interface it is indistinguishable from the other four plants.
Proves
The most common objection is “my machines can’t do that”. They don’t have to. What they have to be able to do is count — and any machine can. What it cannot do is stated there too: this plant has no intervention button, because there is no write path.
- 04
Trigger a fault
3 minShow
The button here triggers a tool breakage at ST020. Then switch to the plant page.
Say
The call goes through the OPC UA method InjectFault — the same way a third-party control system would have to go. Watch the buffer in front fill up and the stations behind starve.
Proves
The layers really are separate. And: buffers decouple — output collapses later than the station does, and that is exactly why the OEE of a line is not a simple multiplication.
- 05
What the control loop makes of it
4 min Go thereShow
Control-loop page, go through all four tabs.
Say
Four plants, one control loop. What differs is the domain knowledge: here the bottleneck, there the queue, there the back-pressure, there the exotherm. Every proposal states its reason, its assumption and the expected effect in numbers.
Proves
The expected effect is fixed beforehand. Without it, it could not be checked afterwards whether the proposal was good — you would have automation without learning.
- 06
Approve a proposal
4 min Go thereShow
Approve an open proposal and watch the trial run.
Say
Approved does not mean adopted. The value is run for a limited time, and afterwards the measured effect decides. The watchdog runs alongside the whole time and aborts without waiting for the end.
Proves
The difficult part is not the intervention but the proof that the forecasts are right. And that takes time: a trial run lasts two to six hours of plant time so that the effect stands out from the noise. At 20× time-lapse that is six to eighteen minutes — start it here and come back to it at step 9.
- 07
The evidence
3 min Go thereShow
Chain of evidence at the bottom of the page, plus the check at the top right.
Say
Every step is listed here: who, what, why, through which gate, with whose approval. Hash-chained and signed. The database rejects UPDATE, DELETE and TRUNCATE — whoever changes an entry breaks the chain from that point on, and the check names the number.
Proves
That is the argument a dashboard cannot make. And the precondition for an operator ever giving the software more freedom.
- 08
Looking back
2 min Go thereShow
Time travel: query the state at any point in time.
Say
No value is ever overwritten. The twin from two hours ago can be queried just like the one from now — and with it the situation on which a decision was based.
Proves
A decision can be reconstructed against exactly the state on which it was made. Without that, an evidence ledger is just a claim with a timestamp.
- 09
Why the optimum is not at the limit
4 min Go thereShow
Learning page, chapter 08, the picture with the two curves.
Say
At the reactor the number of batches keeps rising up to 92 degrees. The number of sellable batches peaks at 88 and collapses afterwards. The same runs, two answers. Further down comes the part nobody else shows: how much effect can be measured at all.
Proves
Whoever optimises the wrong KPI drives into an emergency shutdown — which is why an optimiser without domain knowledge of the plant is dangerous. And whoever claims an effect smaller than the noise of their KPI claims nothing at all. The loop then says “indistinguishable” instead of “confirmed”.
What to expect
- “Does it really run?” — Yes. The plant runs at 20× speed so that a shift can be shown in a quarter of an hour. Everything else is real: protocols, database, twin, evidence.
- “Why is the loop proposing nothing right now?” — Because the situation does not warrant it. A loop that adjusts something at every opportunity is not one. The control-loop page gives the reason in plain language.
- “And if the AI talks nonsense?” — It may do nothing without approval. After that it runs for a limited time, is measured, and goes back if the effect fails to appear. Autonomy is off, and the policy states what would have to happen for it to be up for debate.
- “Why does it say for the cell that nothing is measurable?” — Because it is true. There the manipulated variable acts on one of five parallel machines, and the effect stays below the natural fluctuation. Writing that down openly is the difference between a measurement and a claim — and the most convincing argument of this demo.