Knowing more numbers isn't comprehension; explaining what the plant is doing with them is

What it is

Comprehension is the second level of situation awareness: combining already-perceived readings, states, and events into an explanation of what the plant is currently doing. Comprehension is not "knowing more numbers" — it is establishing the causal relationships between numbers and their bearing on goals. A rising pressure reading is a first-level fact; recognising that it combines with a closed downstream valve into a blockage risk is what this layer produces.

Why it happens

Comprehension runs on a mental model that maps otherwise independent readings onto causal structure, goals, and constraints. A display that only lists discrete values leaves that mapping entirely to working memory, which has limited capacity — once an operator must track much more than four or five variables at once, integration starts dropping detail or gets replaced by a cruder heuristic. A display that shows the relationship directly, such as plotting pressure and downstream valve position on the same trend, or that marks the normal operating envelope, offloads that integration onto an external representation; the operator only has to spot the deviation, not reconstruct the causal chain from scratch.

There is a condition here that can flip the conclusion. Experienced operators typically pattern-match faster, because common fault signatures have been compressed into templates they can match directly. But the same experience produces confirmation bias: once a familiar template is matched, subsequent evidence tends to get interpreted to fit it, even after the real situation has drifted away from that template. Fast comprehension is not the same as correct comprehension — template-matching speed and diagnostic accuracy diverge exactly when evidence is ambiguous, and whether the display can surface data that contradicts the current hypothesis determines whether that divergence gets caught in time.

Studying it

After freezing a scenario, ask why the system is in its current state and which goals are threatened, scoring explanation accuracy and diagnostic consistency. Process tracing asks operators to verbalise how their hypothesis updates as new evidence arrives, recording when and in what direction the revision happens, rather than only scoring whether the final conclusion was right. Question design must avoid anything that reduces to recall — "what is the current pressure" measures the first level, not this one; a genuine probe asks about a causal relationship or a predicted consequence.

A useful design deliberately includes misleading automated advice: the same set of readings is paired with a correct automated diagnosis for one group and a plausible-but-wrong one for another, then the resulting explanations are compared, along with whether operators revise or simply carry forward the original conclusion once contradicting evidence appears. This separates an operator's own comprehension from acceptance of a system's conclusion — the latter effectively outsources the comprehension step.

Where it stops holding

Reciting a display label correctly is not evidence of comprehension. An operator can read out "temperature 85, pressure 3.2 MPa" without being able to say how the two relate; the error there sits in integration, not perception, and the fix is different — improve the relational display, not the font size. Experts can also form different but individually defensible explanations from incomplete evidence; that is not a failure, it means the evidence itself does not yet determine a single cause, and the right response is to flag the evidence gap rather than force one diagnosis.

Automated advice systematically reshapes the operator's hypothesis space: once a system states a conclusion, operators tend to interpret later evidence to fit it, even when that evidence points elsewhere. Operators who work this way for long enough show early signs of the out-of-the-loop performance problem — the capacity to build one's own causal model atrophies with reliance, so that when independent judgment is actually needed (a system fault, a doubtful conclusion), recovering comprehension takes longer than for an operator who was never that dependent. The practical consequence is that comparing display designs requires scoring an operator's own comprehension separately from acceptance of the system's conclusion — the two look identical while automation works, and only diverge once it fails.

Applying it

Co-locate related variables, setpoints, constraints, and recent control actions in the overview, with a direct route from an anomaly back to its causal context, instead of asking operators to reassemble numbers scattered across pages from memory.

How to check: use unprompted verbalisation — no leading questions, just have the operator say freely what is happening and why, then test whether that explanation predicts the evidence that arrives over the next few minutes. Only an explanation that anticipates later evidence counts as real comprehension; a correct action taken by luck does not. For displays that lean on automated advice, add a variant that hides whether a suggestion came from the system or the operator, and compare how much scrutiny and verification each version draws.

Related

• Same group: Y1.01.1 Perception of elements · Y1.01.3 Projection of future status · Y1.01.4 Breakdowns across situation-awareness levels
• Nearby: Y3.02 Mode confusion and accidents · Y1.07 Shift handover
• Search terms: situation comprehension · mental model · automation bias · out-of-the-loop performance problem