A monitoring pane can be accurate and still mislead you. Ours showed three bare numbers for a router's load and, a few lines below, a different number labelled "5-minute load." Both came from real reads. Neither was wrong. Understanding why turned into a small change in how the pane prints its numbers.

Three numbers with no names

The first problem was simple. The router line printed load 0.15 0.20 0.25, three numbers side by side with no indication of what they meant. Anyone who knows the convention can guess they are load averages over 1, 5 and 15 minutes. That guess is exactly the kind of knowledge a pane should not ask a reader to supply at a glance, especially when the screen is meant to be read by whoever happens to come to it next.

The line now prints those numbers as load 0.15 0.20 0.25 (1/5/15 min). Nothing about the measurement changed. Only the label did. A bare triple was a puzzle; a labelled triple is a reading.

The two five-minute numbers

The second problem was more interesting. A different line in the same pane also reported a 5-minute load, and it did not always match the middle number on the router line. That looks like a contradiction: two measurements of one quantity disagreeing on one screen.

They are not the same measurement. The router line comes from a live read cached for a minute; the other line comes from a recording made every five minutes. Each carries the age of its own read. Two figures called "5-minute load" can therefore differ honestly: they are two moments, sampled by two paths, not two answers to one question. The change prints each read's age beside it and explains the difference on the page, so the numbers stop looking like a disagreement and start looking like what they are.

The general lesson is narrower than "numbers can differ." Before treating two figures as a contradiction, find out whether they were sampled at the same moment by the same path. If they were not, the mismatch is information about timing, not an error to be hunted down.

A failed read must not look calm

The third change was the one that mattered most. The router read can fail: a dead path, a malformed reply. The old code parsed the reply and formatted whatever it got. If the load field was missing or garbled, it could still produce a tidy-looking line, because formatting a bad value is easy once you have stopped checking it.

A monitoring line that looks normal when its read has failed is worse than no line, because it invites confidence. The read now validates the shape of the load field first. If it is not three load numbers, or a full five-field load line, the pane prints UNKNOWN with the age of the attempt instead of a plausible-looking value. The short cache still holds on failure, so a dead path is not retried on every redraw.

That is the same rule the rest of this system uses: missing, stale or broken evidence stays visible as missing. It is never rounded into a reassuring sample.

The habit

When you put a measurement in front of a reader, three questions are worth asking. Do the numbers say what they measure? If two figures share a name, were they taken at the same moment by the same path? And when the read fails, does the pane say so, or does it quietly print something that looks fine?

All three fixes here were small: a label, an age and an honest failure case. None changed what was measured. Together they changed whether a reader could trust what the pane appeared to say.