Jon Moshier / Notes / Recognition-Primed Decision Model budding
Note · From the Notebook

Recognition-Primed Decision Model

How experts decide under time pressure without comparing options: the pattern-matching plus mental-simulation model Gary Klein built from fieldwork, and where it breaks.

The Recognition-Primed Decision model (RPD) describes how experienced people decide when there is no time to weigh options. They do not generate a list and score it. They recognize the situation as a type they have seen before, the type comes pre-loaded with an action, and they run that one action forward in their head to check it. Gary Klein built the model from interviews in the field, not experiments in a lab, which is why its claims are about firefighters and nurses rather than undergraduates. It is the mechanism behind one paragraph of my Decision Frameworks note, opened up.

The fireground study that started it

In the mid-1980s Klein, Roberta Calderwood, and Anne Clinton-Cirocco interviewed 26 fire ground commanders, mean experience around 23 years, and probed 156 decision points from nonroutine incidents. The expectation, from classical decision theory, was that commanders would generate options and compare them. They did not. In fewer than 12% of decision points was there any sign of two or more options being compared. In over 80% the commander read the situation as typical and took the action that was typical for it.

The sharper finding was what the commanders said about themselves. They denied making decisions at all. They reported that they simply knew what to do, and many resisted the idea that there had been a choice to make. This is the work that launched Naturalistic Decision Making ([private link]) as a field, funded by a U.S. Army Research Institute project on how people decide in the wild rather than at a desk.

What recognition hands you

The core claim is that recognizing a situation is not just labeling it. Recognition delivers four things at once: the cues worth attending to, the expectancies about what should happen next, a set of plausible goals, and a typical action. The action arrives bundled with the diagnosis. That is why the commander experiences no gap between seeing and doing.

Klein’s most-told case shows recognition by way of its failure. A lieutenant led a crew into a house to fight what read as a routine kitchen fire. The water did not knock it down. The room was hotter and quieter than a fire that size should be. None of this was reasoned through. The expectancies the situation came with were being violated one by one, the violations registered as a feeling that something was wrong, and the lieutenant ordered everyone out. The floor collapsed seconds later. The fire was in the basement underneath them, which is why the living room was hot from below and why the water aimed at the kitchen did nothing. The “sixth sense” was expectancy mismatch firing below the level of words.

Mental simulation and satisficing

Recognition supplies one candidate action. The second stage tests it. The decision-maker runs the action forward as a mental simulation, imagining how it would play out, and acts if it holds. If the simulation surfaces a problem, they adjust the action or, only then, reach for the next option. Options are evaluated one at a time, in sequence, never side by side.

This is satisficing, Herbert Simon’s term ([private link]): take the first option that works, not the best option available. Klein’s commanders were not finding optimal moves and did not try to. Because their experience made the first generated option usually a workable one, sequential satisficing cost them little and bought them speed. The trade the model names is explicit. You give up any guarantee of the best action to get an adequate action now. Under a collapsing roof that is the correct trade.

Three variations

Klein’s model is not one path but three, scaling with how strange the situation is.

The variations matter because they show RPD is not “intuition instead of analysis.” Mental simulation is deliberate, effortful, and conscious. The model puts a fast recognition step in front of a slow checking step, rather than replacing one with the other.

Where it breaks

The same mechanism that catches the basement fire can manufacture a false read. Mental simulation runs on the scenario recognition already proposed, and the simulation tends to defend that scenario. Research on the technique found people distorted incoming data to fit the story they had already built, an expectancy bias where you see what the pattern says you should see. A wrong recognition does not announce itself; it feels exactly like a right one.

So RPD is trustworthy only under conditions. Klein and Kahneman, who started from opposite priors about intuition, mapped the boundary together in “Conditions for Intuitive Expertise: A Failure to Disagree” (2009). Skilled intuition is real, but only in a high-validity environment that holds stable, learnable patterns and returns fast, clear feedback. Firefighters and anesthesiologists get both. Stock pickers and long-range forecasters get neither, and there the confident pattern-match is an illusion. RPD also degrades with an inexperienced operator or a genuinely novel situation, exactly the cases where pattern-matching has nothing valid to match against. In Cynefin terms it is a tool for the Clear and Complicated domains, and a hazard in the Complex one.

There is a standing methodological objection too. The evidence is built from retrospective critical-incident interviews, and the central recognition step is tacit, so it resists the kind of test that would distinguish real pattern-matching from after-the-fact narration. The model is descriptively rich and hard to falsify in the same breath.

Try it

Tag your own decisions for a week (15 min/day, a notes file). For each non-trivial call, write one line: was it instant recognition or did you compare options, what cues triggered it, and what you expected to happen next. At week’s end, sort by your own expertise. In your strong domain the recognition rate should run high; in a domain you are new to, you will find yourself comparing options or guessing. The split, in your own log, is the high-validity boundary drawn on your life. Watch the expectancy column: the entries where what you expected did not happen are your real-time check on whether the recognition was sound or just confident.

Run a single-option premortem (20 min). Take a decision where one answer already feels obvious, the recognition has fired. Before acting, force the mental-simulation step the model says experts run: narrate how this action plays out, then deliberately hunt for the cue that would mean you misread the situation. You are trying to feel the pull described in the research, the urge to explain the contradicting cue away. Noticing that pull is the difference between checking a recognition and falling in love with it.

See also

Sources

← All notes Read recent essays →