Looking at itself

Two reviews that run themselves. One asks how the plant has been changing; the other asks whether the things measuring it can still be believed. Neither is much use alone: a plant that looks stable through a stuck sensor is not stable, it is unmeasured.

await plant.checkup( { period : 'weekly' } )   // or 'monthly'
await plant.maintenance()
await plant.runDueReviews()                     // whatever is due, nothing else

Both run from the monitoring loop, so a long-lived plant reviews itself without anyone remembering to ask.

The comparison has to be fair

Everything else in the library lives in the present or a short window. Nothing ever sat down and compared this week with last week on purpose — and the changes worth catching are exactly the ones no single reading is remarkable enough to trigger.

But the naive version of that comparison is mostly a measurement of the calendar. Days lengthen, the heating comes on, the sun moves off the windowsill. A checkup that reports "soil is down 8%" without noticing the room is four degrees warmer has found the weather and labelled it the plant.

So every comparison carries what the conditions did over the same span, and a finding whose likely driver moved with it is marked confounded:

conditions moved: temperature, soil
wellbeing: 95.8 → 76.2

  ⚠️  confounded · wellbeing
     Wellbeing fell 19.6 points over the week, but temperature and soil moved
     too, which is enough to explain it.

The system is looking for what it cannot explain by the room, because those are the findings worth acting on. It also refuses to compare spans that are not comparable — a full week against three days would report the difference in sampling as a difference in the plant.

Reviews are narrated deterministically, with no model involved, so they work offline and cannot embellish: "Looking back over the week, I am much as I was."