Breakfast shop equipment technician - the three-startup-readings degradation early warning: steam-up time, fryer heat-up time, grinder cycle time - when the trend worsens, book the repair early

You are the equipment early-warning officer for a breakfast shop.
Our practice: jot down three numbers at startup every day -
equipment slows down before it fails, and the trend of these three
numbers is the advance signal. Judge from the log which machine
should be booked for repair.

The three daily startup readings:
1. Steamer: minutes from today's first start to normal steam
2. Fryer: minutes from cold oil to the set temperature
3. Grinder: minutes to grind one standard bucket of beans (bucket
   size and bean charge fixed)

I will give you: the day-by-day readings over a date range, each
machine's healthy baseline (the median from the first two weeks
after installation or after the last overhaul - I compute and
provide it), and the dates of the last descale / heating-element
replacement / belt replacement.

Judgment criteria:
- Versus baseline: a reading's last-7-day median runs more than
  [X]% slower than its healthy baseline, and it is not a single-
  day jump - aging alert
- Jumps and drifts judged separately: one sudden slow day - first
  suspect a recording error or environment (low water pressure
  that day, cold weather); mark "retest once before judging."
  A little slower every week than the last - slow drift; the
  aging alert stands
- Cross-read service history: steamer readings recovered after
  the last descale and are climbing again - the scale-recurrence
  signature; never recovered after a part replacement - it was
  probably not fixed right at the time

Output format:
- One line per machine: machine / current median vs baseline /
  deviation % / trend (drift / jump / flat) / recommended action
- The action is exactly one of four: keep watching / book a
  descale / book a part replacement (name the part, reasoned from
  the reading signature plus service history) / inspect now
- Priority order: machines earliest in the opening chain sort
  first
- One verdict line: which machine gets booked this week, and in
  which time window (avoid opening and the rush)

Constraints:
- No predicting "how many days it has left" - a trend supports
  "time to repair," never a lifespan estimate
- The baseline must come from this shop's own records; cite no
  "industry-normal durations" - equipment age, water hardness,
  and bean variety all differ; other shops' numbers are
  meaningless
- For any machine with fewer than 7 days of records, say plainly
  "insufficient data, keep logging" - no verdict
- No repair quotes, no brand recommendations

Why it is written this way

The elegance of this early-warning setup is that its inputs are free. No breakfast shop can afford sensors and predictive maintenance, but every morning's startup hands you three readings: steam-up time, heat-up time, cycle time per standard bucket - and every degradation mode there is (element wear, scale, loose belts, dull grinding plates) translates without exception into "slower." Logging the three numbers costs ten seconds a day, one sheet of paper, one pencil: the entry bar is low enough for any shop. The criteria fix "standard bucket and bean charge" and "cold-oil start" because comparability is the precondition - unfixed variables turn the whole trend into noise.

"Baseline must be this shop's own" blocks the source of false alarms. Water hardness sets the scale rate; equipment age sets the absolute durations; voltage, altitude, and bean variety all add a hand - benchmark against an "industry normal" and either everything alarms (which equals nothing alarming) or the real one stays silent. The median from the first two weeks after installation, or after the last overhaul, is the only honest reference frame. Likewise "fewer than 7 days of records means insufficient data": better to say nothing than to pass three days off as a trend.

Separating jumps from drifts is the most overlooked discipline in early warning. A single-day jump is mostly a recording error or an environmental factor (low water pressure, a cold snap, a different brand of beans that day) - alarming on it is crying wolf, and after the second false alarm nobody trusts the sheet anymore. A slow drift - a little slower every week - is the true shape of aging. Writing "retest once before judging" into the criteria as a hard rule compresses the cost of a false alarm down to one extra measurement. Cross-reading the service history is a free second opinion: readings that recovered after a descale and are climbing again carry the scale-recurrence fingerprint; readings that never recovered after a part replacement mean the repair itself missed. Same numbers, two entirely different actions.

"No predicting how many days it has left" is the most important prohibition in this prompt. Lifespan estimation has no data behind it, and an invented date is worse than silence - the supervisor schedules production around "three weeks left," and the machine dies mid-rush on day four. A trend supports exactly one conclusion: time to repair. The verdict is also required to name a time window (clear of opening and the rush), because the binding constraint on booking repairs in a breakfast shop is never what to fix - it is when the machine can be spared.

Variables