Chapter 1 - The Signal

The number had been on the screen for seven minutes.

41.

Simon Keller had stopped checking his watch.

“Forty-one cases,” he said.

Marek nodded.

No one else at the table reacted.

Beyond the windows of the conference room, Zurich sat beneath a January sky that seemed unable to decide whether to become light or dark. Far below, trams moved through the intersection. From up here, their paths looked inevitable.

Simon turned back to the screen.

The presentation had stopped on slide seven. No headline, just two columns and a chart.

Total population: 48,214

Segment X: 41

Below them, two bars.

1.1 percent.

4.9 percent.

“And you're saying that's enough?”

“I'm saying it's enough not to pretend we haven't seen it.”

Marek sat diagonally across from him. His laptop was closed. He had barely looked at it since the meeting began.

Martin Reiner leaned back.

“No one's doing that.”

Marek looked at him.

“Yes, we are.”

Reiner let the answer sit for a moment.

Simon knew this kind of silence. It wasn't uncomfortable. It was useful. People often corrected themselves inside it.

Marek didn't.

Simon pointed at the chart.

“Two cases.”

“Two clear misroutings.”

“Out of forty-one.”

“Out of forty-one cases with the same combination of features.”

“That's the problem with a small sample.”

“No,” Marek said. “The problem is that we don't know whether it's small, or whether we just don't look often enough.”

Reiner leaned forward for the first time.

“What are you proposing?”

“Take the segment out.”

“Of the rollout?”

“Of the automated pathway.”

Simon looked at Marek.

“For how long?”

“Until we know more.”

“That could take a year.”

“Then it takes a year.”

Something shifted in Reiner's face. Not quite a smile.

“So we delay part of the programme for a year because of two cases.”

“No.”

Marek stood and walked to the screen.

He pointed to the second number.

“Because of 4.9 percent.”

Simon shook his head.

“You know you can't present it like that.”

“Why? That's the number.”

“Because you're comparing a rate based on forty-one observations with a population of nearly fifty thousand.”

“So we ignore it?”

“We treat it for what it is.”

Marek waited.

Simon said, “A signal.”

For a moment, the room was silent.

“Exactly,” Marek said.

Simon knew he hadn't won anything with that.

On the table between them lay the rollout documents. Eight months of work. Three external reviews. Two validation rounds. A pilot that had performed better than even Simon had expected.

PROTOKOLL wasn't spectacular.

That was one of the reasons he trusted it.

It didn't make diagnoses. It didn't decide which treatment a patient received. It didn't replace a doctor.

It sorted.

Most cases continued through the normal process. Clear cases could move faster. Complex or unusual combinations were routed to people: case management, medical review, further assessment.

The principle was mundane.

The effect wasn't.

For years, the number of cases had grown faster than the number of people available to review them. If you tried to check everything, eventually you checked nothing properly.

PROTOKOLL was supposed to help decide where human attention was needed most.

Which was exactly why Marek's slide was uncomfortable.

Simon pulled the printed validation summary towards him.

“Overall accuracy?”

“Ninety-one point four.”

“Expected?”

“Ninety-two.”

“False negatives?”

“One point one.”

“Compared with the existing process?”

Marek didn't answer immediately.

Simon looked up.

“Marek.”

“Worse.”

Reiner looked from one man to the other.

“By how much?”

“The existing process is around one point eight.”

Reiner raised an eyebrow.

Simon said nothing.

That was the number that carried more weight in this room than the others.

Not because it disproved them.

Because it made them harder to interpret.

With PROTOKOLL, fewer cases would incorrectly remain in the standard pathway than they did today.

Not theoretically.

In the data.

Simon turned a page.

“So we're introducing a system that reduces the relevant error rate overall from one point eight to one point one.”

Marek sat down again.

“Yes.”

“And you want to change the rollout because we're seeing two misroutings in a group of forty-one cases.”

“Yes.”

“Even though we don't know whether the effect is stable.”

“Yes.”

Simon put the papers down.

“Why?”

Marek looked at him as though Simon had deliberately asked the wrong question.

“Because models can look good on average and still fail in exactly the places where we need them.”

“I know that.”

“No.”

Reiner raised a hand.

“Marek.”

But Marek kept looking at Simon.

“You know it's theoretically possible. That's not the same thing.”

Simon felt irritation. Not much. Just enough to register it.

He had learned to treat irritation like a number: notice it, classify it, don't act on it.

“Then explain the difference.”

Marek opened his laptop.

“The two cases don't just share the same routing category. They share four features. Young age. Unusual treatment history. High expected cost. Low model confidence just above our threshold.”

Simon looked at the chart.

“And?”

“And there are other cases where the decision was technically correct, but confidence becomes unstable in the same place.”

“How many?”

“Five.”

“Misroutings?”

“No.”

“Then they're not misroutings.”

“Not yet.”

Reiner exhaled audibly.

“I can't take that to the Steering Committee.”

“Neither can I,” Marek said.

Simon looked at him.

Marek closed his laptop.

“That's why we're sitting here.”