A monitoring map has two surfaces. One is a live pane that a person opens to see the current state. The other is a saved text twin written next to it, so a reader without the pane can see the same thing later. Both are supposed to describe the same collector. In practice one always printed "phone history UNKNOWN (no usable collector rows)" while the live pane showed a derived line from the same healthy source. Nothing was wrong with the data. One of the two readers was broken, and the wording made that indistinguishable from an empty feed.
The argument that was never passed
The saved view builds its text by calling the map's model builder. The live pane builds the same model through a different path. That is a reasonable arrangement until the two paths stop passing the same arguments.
The builder takes a stability value: a derived summary of recent signal history, computed from the collector's rows. The live path computed it and passed it in. The saved path passed None in its place. The collector file was healthy and full in both cases; the saved reader simply never looked at the history, so it printed the fail-closed message on every render.
This is the sort of bug that survives review because each half looks correct in isolation. "If there is no stability input, say UNKNOWN" is the right rule. "No stability input" was also literally true in the saved path. The defect was upstream of the rule: the input had been dropped at the call site, not missing from the source.
UNKNOWN can be a false negative too
There is a useful rule that a missing input must produce UNKNOWN rather than zero. It keeps a report from inventing a measurement. This case is the mirror image, and it is worth stating plainly: UNKNOWN is honest only when the input really is absent. When the input exists and a reader path discards it, UNKNOWN becomes a false negative. It hides data that is sitting right there.
For a reader, the two cases look identical. Both print the same three words. The live pane next door is the only clue, and you have to know to compare them. A saved artefact has no such neighbour — it is the entire evidence a later reader gets.
The practical distinction is not "data or no data". It is "no data" versus "no working reader for the data". A system that cannot make that distinction will sometimes report a healthy input as a broken one, and the report will be trusted because it used the careful word.
One read, two callers
The repair was to stop computing this in two places. A single helper now reads the collector file once and returns both the newest row and the derived stability history. The live pane calls it. The saved twin calls it. There is no longer a call site that can pass None while the other passes a real value, because neither path constructs the argument by hand any more.
The fail-closed behaviour stays. If the collector file is genuinely missing, the helper returns UNKNOWN and the saved view still says so. Two tests pin both sides: one asserts that the saved twin shows the same derived line as the live pane when the collector is healthy, the other asserts that it still prints "phone history UNKNOWN" when the collector is absent. In the fixture built for the first test, three synthetic samples turn into a short derived sentence about recent drops and freshness; the point is not the numbers but that both surfaces now derive them from one read.
The rule that travels
If two surfaces claim to show the same thing, make them call the same read. Parallel implementations of one observation drift, and they drift silently — especially when one of them fails in the cautious direction. A message like UNKNOWN is a signal of care only when you can trust that every reader reached the source.
The smaller habit is to ask, when a pane says "no data", whether the data was missing or the argument was. The two have the same output and opposite meanings. Only one of them is a fact about the world.

