Reports
Alarm health, recurring faults, the pilot measurements and the evaluation report.
Reports shows two reports built from the alarm history, the stops and the confirmed diagnoses. Every role can open it.
Alarm health
For the last 7 days by default (ISA-18.2 and EEMUA 191 measures):
- Alarm rate: activations per hour;
- Time in flood: the share of the period in an alarm flood, and the number of floods;
- Most frequent alarms;
- Chattering alarms;
- Standing more than 24 hours: alarms active for over a day, with when they started and cleared.
Recurring faults
For the last 30 days by default, the stops grouped by cause, ranked By number of stops and By time down. Each row gives the number of stops, minutes down and the Confirmed diagnoses; a stop not yet restarted is counted separately. A period with no confirmed cause says "No confirmed causes in the period."
Pilot measurements
The pilot measurements (median time to diagnosis against the site's week-0 baseline, the share of answers rated useful, how often the top cause was the confirmed one, and sessions per shift) are computed from the feedback and served at GET /api/kpi?start=…&end=…. The baseline and the shifts are set in site.yaml.
Planned. A KPI panel on the Reports screen. Today the measurements are read from the API.
Evaluation report
Before go-live, the site's own past stops are replayed and scored with python -m eval.site_replay cases.yaml (top-1 and top-3 against the known causes). See Run the week-0 assessment.
4M also evaluates BooleX every night on the fault library of the bench and the sector demo plants, headless on recorded data: for each model, how often the top cause and the top three causes hold the right one, and the share of expected citations given. A run fails when a model falls more than 5 points below its committed baseline, so a change to the prompt, retrieval or model that makes answers worse is caught before it is released. Each run also counts each model's input and output tokens per answer and, where a price is set, an estimated cost per answer and per run, to help a site choose a model. Ask your 4M contact for the latest report for the model you run.