An insight card is a small claim with its evidence attached, not a verdict handed down from a model nobody can see inside. Every field on it means something specific, and once you know what each one is actually saying, a card takes seconds to read instead of a guess at whether to trust it.
The anatomy of a card
Title and severity. The title states the finding in plain words, running below pattern, drifting off band, precedes a stop. Severity, info, warning, or critical, reflects how far off normal the reading sits, not how the finding was produced.
Detail. A sentence restating the evidence in plain language, expected against observed, or a metric’s pre-event average against its baseline. It’s the same numbers the evidence carries, written for a person reading quickly instead of parsing a data structure.
Window and driver. Which stretch of time the finding is based on, and which specific comparison drove it, a hour-of-day baseline, a control band, a precursor’s pre-event mean. This is what makes a card checkable instead of a bare assertion: the exact basis is named, not hidden behind the headline number.
Confidence. A value between zero and one that grows with how much data backs the finding, more samples, more support, closer to one. A card with a low confidence value is telling you the same thing a wide error bar would: this comparison is real, but it’s standing on less history than a fully mature one would be.
Dollars, or a dash. A shortfall in availability or rate, and a recurring downtime reason, get priced against the machine’s resolved cost rate, and the card names that rate’s source. A raw sensor drifting off its band, and a precursor pattern, show no dollar figure at all, because turning either into a price would mean inventing a relationship between a reading and a cost that isn’t actually there. A dash here isn’t a gap in the product, it’s the product refusing to make up a number.
Proposed rule. Some cards, mostly precursor findings, carry a one-click offer to turn the pattern into a live trigger, a threshold or an alarm-repeat rule, prefilled from the exact numbers the pattern showed. Not every card has one. A card without this section simply didn’t earn an actionable trigger from its own evidence.
What “learning” and a withdrawn card are actually saying
A machine, or a specific signal on it, that hasn’t cleared its sample floor yet shows a learning state instead of a card. That’s not an empty screen standing in for a bug, it’s the platform saying it has too little history to make a comparison it would stand behind. The same discipline runs all the way up to the failure-risk estimate on the Maintenance page: it predicts a failure from a machine’s own count of past breakdowns and shows the evidence behind the number, and it says not enough history when that count sits below what a credible estimate needs, instead of showing a percentage stretched over almost nothing.
A card can also disappear on its own, without anyone dismissing it. If the signal behind it later gets reclassified, a counter that was mistakenly treated as a continuous reading, a tag that finally got a real purpose set, the model built on it retires and any open card tied to it withdraws automatically, with the specific reason recorded. That’s a correction, not a dismissal, and it’s why a card that seemed to vanish overnight usually has a clear explanation waiting on the same machine’s history if you look for it.
Reading the age of a card
An open card doesn’t rewrite its own numbers every five minutes just because scoring runs that often. Its title and detail freeze at the moment it first fired, so it can stay open for days without its wording changing, even though the condition behind it may still hold. Once a windowed card has sat open for more than twice as long as the window it was measured over, it starts showing how long it’s been flagged, so a reading from three days ago never reads as something that just happened. A pattern mined from history rather than a live reading, a precursor, carries no such age label at all, since it was never a claim about right now in the first place.
A borrowed card reads differently, and says so
A brand new machine can show findings within a day even before it has its own learned pattern, when a similar machine on the same floor already has one. A card built that way names the fact directly in its own wording, something like “running far less than its learned pattern, borrowed from CNC #1 until this machine has a day of history.” That suffix is doing real work: it’s telling you the comparison behind the card came from a different machine’s history, not this one’s, so the finding is a reasonable starting estimate instead of a fully earned one. The moment the new machine’s own data clears its floor, later cards on that same kind of finding drop the suffix on their own, no dismissal needed, because the comparison is finally standing on the machine’s own history.
When more than one machine shows the same pattern
A single card can also represent more than one machine at once. When the identical kind of finding, the same drift, the same recurring downtime risk, shows up on several assets, the Opportunities feed collapses them into one card listing every machine it applies to, sorted by dollar impact, instead of a wall of near-duplicate cards saying the same thing five times. Clicking into any one machine’s own detail still shows its own evidence separately, its own window, its own driver, its own confidence, so the collapsed view is a convenience for scanning, not a blending of five machines’ numbers into one.
Explain, dismiss, and what each one does
Clicking Explain sends a card’s own evidence to the AI layer for a plain-language elaboration, it doesn’t compute anything new, it restates what’s already on the card in fuller sentences. Dismissing a card marks it addressed and frees that exact combination of asset, kind, and model to fire again if the same pattern shows up later, it doesn’t delete the finding’s history, which stays visible if you go looking for it.
Quick recap
- A title and severity state the claim. The window, driver, and detail state exactly what comparison produced it, so nothing on the card is unbacked.
- Confidence grows with sample size, a low value means the comparison is real but standing on thin history, not that the finding is wrong.
- A dollar figure only appears where the pricing math is real. A dash means the finding has no defensible price to show, not that the product skipped a step.
- A proposed rule appears only on patterns that earned one, prefilled from the data that produced them, never typed from memory.
- A learning state and a withdrawn card both mean the same underlying thing: not enough history, or history that changed the classification, stated plainly instead of guessed past.