Placeholder copy · the text on this page is a draft

White paper

The judgment loop: what two centuries of automation left to people

Every industrial revolution since 1788 has handed one more control loop to a machine. This paper traces the one that never moved, and sets out what a system would need before a plant could trust it with that call.

Name Surname

Co-founder, Ferry

The problem

An alarm nobody trusted

Reactor 3 had been raising a high-pressure alarm since the spring of the previous year. It fired perhaps twice a week, never in a pattern anyone could pin down, and never when the pressure on the adjacent transmitter was near the trip point.

By the time Ferry arrived the alarm had been through two investigations. Both ended the same way: the instrument checked out, the loop tested clean, and the trip kept firing. The shift team had stopped treating it as information.

What made it expensive was not the alarm itself. It was that Reactor 3 had, over eleven months, trained its operators to discount one of its own safety signals. Every time the trip fired and nothing was wrong, the cost of ignoring it went down slightly.

The site knew this. It appears in two shift reports and in the minutes of a process safety meeting. Nobody had the hours to take the loop apart while the unit was running, and the next turnaround was eighteen months out.

Key findings

· The judgment loop depends on reading several systems together, which no single plant system was built to do.

· Trust is earned per site and per question. A general claim of accuracy is not enough for a safety case.

· On three live sites, a reasoning system closed investigations that had stalled, without writing to any control system.

The decommissioned transmitter is the highlighted component.

Fig. 1

The last loop

What the last loop actually involves

Consider a trip that fires twice a week and never in a pattern. The instrument checks out. The loop tests clean. The shift team stops treating it as information. Somebody has to decide what is really happening, and that decision draws on the logic as configured, four years of history, the change records and the procedure the trip was meant to protect.

An experienced engineer does this by holding all of those in view at once. It takes days, and most sites do not have the days. So the alarm stays, and the plant slowly learns to ignore one of its own signals.

“We knew it was wrong. Proving it, in a way the safety case would accept, was the problem.”

Site process engineering lead, specialty chemicals

The judgment loop is therefore less a single decision than an act of reading. Whatever closes it has to read the same sources an engineer would, in the same combination, and show its reasoning in a form the engineer can check.

What we did

Reading the logic, not the documentation

Ferry connected to the historian and to the control system, and built one workflow against one question: what conditions are actually present when this trip fires.

The workflow read four years of tag history alongside the control logic itself. The documentation described a loop with one transmitter. The logic still referenced two. The second had been removed during a 2024 turnaround and its tag was never deleted.

Nothing was taken offline for any of this. The workflow read from systems that were already recording, over a path agreed with the control system owner before anything was connected. No operator was asked to change how they worked while it ran.

The first credible answer arrived inside a day. The remaining three were spent checking it: confirming the removal against the turnaround records, replaying eleven months of firings against the floating signal, and putting it in a form the safety case would accept.

A control room during a shift change. The evidence for most open questions is already on these screens, in separate systems.

Fig. 3

Conditions

Four conditions before a plant can trust it

Every earlier handover met the same tests before anyone relied on it. We think the judgment loop has to meet four.

Condition 1

It reads before it writes

The system reads plant systems and never writes to them. A person takes every action, and the site can close the connection at any time.

Condition 2

It shows its working

Every answer arrives with the sources it used, so an engineer can check it in minutes and sign it.

Condition 3

It is tested per site

Trust is earned on the plant in question, against questions that plant has already answered by hand.

Condition 4

It keeps the site's knowledge

What a plant teaches the system stays with that plant, in its own tenancy, and does not train anyone else's model.

The decommissioned transmitter is the highlighted component.

Fig. 1

What the workflow read

Three sources, one question

Tag history

Four years of process data from the historian, sampled at the rate the plant actually records rather than the rate the trend screen displays. Enough to see whether the conditions the trip claims to detect were ever present.

Control logic

The logic as configured on the system, not as described in the documentation. This is where the case turned: the documentation described one transmitter on the loop, the logic still referenced two.

Change records

Turnaround and management of change records from the last five years, which placed the removal of the second transmitter in 2024 and showed its tag was never retired with it.

“We knew it was wrong. Proving it, in a way the safety case would accept, was the problem. Four days of somebody else's reasoning did what two of our own investigations could not.”

↳ Site process engineering lead, specialty chemicals site, Teesside

The results

And what it led to

The tag was removed from the trip logic under the site’s management of change process. The alarm has not fired since. The same reasoning was then pointed at the rest of the unit: three hundred and forty alarms had accumulated since commissioning, and that review is now a standing job rather than a project that stalls.

There is a second result that matters more to the site than the first. The review that came out of this is not a project with an end date. It runs against the logic as it stands, so when the plant changes the reasoning runs again rather than expiring.

The results

And what it led to

The tag was removed from the trip logic under the site’s management of change process. The alarm has not fired since. The same reasoning was then pointed at the rest of the unit: three hundred and forty alarms had accumulated since commissioning, and that review is now a standing job rather than a project that stalls.

There is a second result that matters more to the site than the first. The review that came out of this is not a project with an end date. It runs against the logic as it stands, so when the plant changes the reasoning runs again rather than expiring.

Limits

What remains open

The tag was removed from the trip logic under the site’s management of change process. The alarm has not fired since. The same reasoning was then pointed at the rest of the unit: three hundred and forty alarms had accumulated since commissioning, and that review is now a standing job rather than a project that stalls.

There is a second result that matters more to the site than the first. The review that came out of this is not a project with an end date. It runs against the logic as it stands, so when the plant changes the reasoning runs again rather than expiring.

A caveat

Three sites is evidence, not proof. The figures here come from pilot engagements, and the sites are not named at their request.

More stories

Food and beverage, Midlands

Four hours of handover, done in forty minutes

Pharmaceuticals, Ireland

Three hundred alarms reviewed in a fortnight

Read the report before you decide

Request the SOC 2 report and we will send it the same day. If you would rather ask an engineer directly, that route is open too.