Most plants can detect. Open the monitoring stack at any refinery or generating station and you will find alerts: a vibration excursion, thermal drift, a pump curve sliding out of position. Detection is the part the industry solved first.
Conversion is the part it has not. An alert that never becomes a work order changes nothing. The honest measure of a reliability program is not how many anomalies it raises, but how many end as completed work with the finding checked against the diagnosis.
Five handoffs sit in between. Loops break at the handoffs, not at the detection.
Alert to diagnosis. An alert says something changed. A diagnosis says what is degrading, by what mechanism, how fast, and with what confidence. Without a mechanism there is nothing to plan against.
Diagnosis to decision. Severity is a property of the signal. Criticality is a property of the plant. A recommendation that cannot express production risk, safety exposure and remaining run time loses to whatever else is competing for the same crew.
Decision to approval. With false positive rates on legacy systems running near 70 percent, skepticism is rational. The evidence has to be legible in minutes: tags, trend, comparable events, recommended action, and what was ruled out.
Approval to work order. The decision has to land in the CMMS as a real object, with the right functional location, priority, craft, parts and task list. Programs stall here when the asset hierarchy in the model was never reconciled with SAP or Maximo.
Execution back into the model. The technician opens the machine. Right, partly right, or wrong? This is where the loop actually closes, and almost nobody instruments it. A confirmed finding sharpens the next diagnosis. A wrong one is worth just as much, provided it comes back.
Measure confirmation rate first, then conversion rate, then cycle time from detection to work order. Alert volume is a vanity metric.
An open loop does not hold steady. Accuracy stalls where deployment left it, engineers stop trusting the alerts, planners stop holding capacity. Most programs that die in year two die here with the models still running.
Closing the loop is not a documentation exercise. It is the commitment that every diagnosis gets an outcome, every outcome is recorded against the diagnosis that predicted it, and that record goes back to whatever made the call. Process reliability compounds from there.