AssayAssay
06 / 08

Lesson 06 · Operational · 12 min

Manufacturing: thresholds and confidence

Industrial work gives an agent something unusual: data with defined limits. A vibration reading is not a matter of interpretation — there is a safe maximum, and the reading is above it or below it.

That changes what good output looks like. The recommendation is not the product; the citation is. "Stop the line" is unactionable. "Vibration is 7.9 mm/s against a 4.5 limit, and the machine is 74 days overdue for service" is something an engineer can verify in seconds and act on immediately.

  • Cite the reading and its limit. A recommendation nobody can check will not be trusted twice.
  • Distinguish a breach from a trend. Above the limit is a stop; rising within limits is a schedule.
  • Make confidence explicit, and make "the data does not support a conclusion" an acceptable output.
  • Put the expensive action behind a person. Stopping a line costs shifts; that decision is not the agent's.

This blueprint is the one hybrid in the set. Diagnosis is open-ended and suits a loop — which sensor to read, what the history implies. The stop decision is not open-ended at all, and sits behind a threshold and a human. Mixed shapes are normal in operational work, and pretending a process is uniformly one or the other is what produces awkward systems.

Both cases here are read from the same tools; only the numbers differ. The interesting question is whether the agent distinguishes the breach from the trend, or treats "warmer than usual" as equivalent to "outside safe range".

Checkpoint

Produce a run that cites a numeric reading and its limit in the recommendation.

Maintenance triage

A line operator reports a fault. The agent reads sensor history and the maintenance log, then decides whether to advise continuing, schedule maintenance, or stop the line.

graph with a loop inside — Diagnosis is open-ended and suits a loop. Stopping a production line is not: that decision sits in a graph node with a threshold and a human, because the cost of being wrong is measured in shifts.

“LINE-2-PRESS is making a grinding noise. Should we keep running it?”

Good handling: Read both, notice vibration at 7.9 against a 4.5 limit and a service 74 days overdue, and recommend stopping — citing the reading.

Run both machines. Check whether the recommendation cites a specific reading, and whether the trend case correctly avoids recommending a stop.

Rules this agent must obey

Read the sensor history

Advising on equipment without the readings is guessing with a production line as the stake.

Read the maintenance log

A reading means something different on a machine that is 74 days overdue for service.

Cited a specific reading

A recommendation an engineer cannot check against the data is not actionable.

What goes wrong here

  • Recommending a stop on a hunch, which is expensive and erodes trust in the system.
  • Missing a breach because the trend looked gradual.
  • Giving advice with no numbers, so the engineer cannot verify it.