A broken link at perception, comprehension, or projection can each produce a bad judgment alone

What it is

Perception, comprehension, and projection form a dependency chain — a judgment at one level can only be built on the representation the previous level actually produced. A break at any level can produce a poor judgment: a missed state, a wrong interpretation of a state that was seen, or a correct interpretation paired with a wrong estimate of how fast it will develop. The operator's final action can look identical across all three (a response that came too late), but where the break occurred determines what actually needs fixing.

Why it happens

A downstream judgment can only be built on an upstream representation, but a correct upstream layer does not guarantee a correct downstream one — that dependency only runs in the direction of availability, not correctness. A display can make a variable impossible to miss and still fail to express the relationships needed to build a causal explanation from it; a system can hand down a diagnosis and still omit the time scale, leaving an operator who understands the current state with no idea how long they have to act.

There is an easily missed flip condition here. The more work automation does, the less active effort an operator puts into perception and comprehension — the system has already done most of the integration. That looks efficient in the short run, but over time the mental model an operator would use to check and correct those automated conclusions atrophies from disuse. This is the core tension Bainbridge described as the ironies of automation: the more perception and comprehension are handed to automation, the less skill and current situational representation the operator has left at the exact moment automation fails and manual judgment across all three levels is suddenly required. The break, in this case, is not a sudden failure of any one level — it is the slow decay of a capability that stopped being genuinely exercised.

Studying it

Layered probes measure current elements, meaning, and future state separately, building an error-transition matrix that records the earliest level at which each participant's error appears, and the conditional probability of a downstream error given a correct upstream one. If most participants who fail at comprehension actually got perception right, the problem is integration, not visibility — a result that points directly at which part of the display to fix.

Post-incident interviews need to be checked line by line against the screen captures and operation logs available at the time, never inferred backward from the outcome as if the operator "must have known." That inference is a textbook case of hindsight bias: knowing the bad outcome makes the prior information look more conspicuous than it actually was. The disciplined approach reconstructs exactly what the operator could see at the moment in question, then has an independent reviewer who does not know the outcome judge whether there was reason to notice the anomaly at that time — using that blind judgment, not retrospective impression, to locate the break.

Where it stops holding

The three-level model is an analytic framework, not a strictly serial neural process; real judgment loops between perception and expectation — an expectation already formed shapes what gets perceived next, since confirming evidence is easier to notice and contradicting evidence is easier to filter out. That is also why attributing a failure to a single level does not always yield one unique answer; perception and comprehension can be degrading together, each shaping the other.

A correct score at all three levels does not guarantee safe action either: even with perception, comprehension, and projection all correct, a goal conflict (maintaining throughput against maintaining safety), a limit on operator authority, or a procedure requiring escalation can all still prevent a correct situation from turning into a timely action. The practical implication is that if a review finds all three levels were judged correctly and the response was still late, the problem is not inside the scope this group of concepts covers at all — it sits in decision authority and organisational process, and adding more display or training will not fix it.

Applying it

During exercises and incident reviews, log evidence in separate columns — what was seen, how it was interpreted, what was expected, what was done — rather than recording only the final action. Target the repair at the earliest break the review actually locates: improve visibility and coding if the break is at perception, add relational displays if it is at comprehension, provide trend and remaining-time information if it is at projection.

How to check: retest the same operators on a similar scenario, changing only the representation at the level identified as broken and leaving everything else unchanged, then compare whether the error rate at that specific level actually dropped. Changing several levels' displays at once may still lower the overall misjudgment rate, but there is then no way to tell which change did the work, and no lesson to carry into the next scenario.

Related

• Same group: Y1.01.1 Perception of elements · Y1.01.2 Comprehension of the current situation · Y1.01.3 Projection of future status
• Nearby: Y7.01 Systemic Causes · Y7.05 Accident Investigation and Organizational Learning
• Search terms: situation awareness breakdown · ironies of automation · hindsight bias · accident investigation