Reactor type is not taxonomy, it is a different control problem
Every reactor on this site is classified by two things: what cools it and what slows its neutrons. It reads like filing. It is not. Those two choices decide what happens in the minutes after something goes wrong, which is why a control room qualified on one type does not transfer to another.
Here is what the classification actually buys you, in the four families you can put hands on.
The coefficient that decides which way power moves
The single most consequential number in a reactor is the void coefficient: what happens to reactivity when coolant boils away.
In a pressurised water reactor, the water is both coolant and moderator. Lose water and you lose moderation, neutrons stay fast, fission falls. Power drops on its own. The physics pushes toward shutdown, and 311 of the 419 operating reactors worldwide are built this way.
In an RBMK, graphite does the moderating and water mostly absorbs. Boil the water at low power and you remove an absorber while the graphite keeps moderating happily. Reactivity goes up. Power rises, which boils more water, which raises power again. That is a positive void coefficient, and it is not an abstraction: it is the mechanism that destroyed Chernobyl 4 during a low-power test. Seven RBMKs are still operating.
CANDU sits awkwardly in between, with a positive void coefficient of its own, answered not by the physics but by engineering: two fully independent, fast shutdown systems, one dropping rods and one injecting poison into the moderator.
Same accident question. Three different answers, and the answer follows directly from the coolant and moderator columns in the table.
What the operator actually reaches for
The control problem differs just as much as the physics.
A PWR operator adjusts boron concentration in the coolant for slow, fleet-wide reactivity changes and saves the control rods for fast ones. Power is set at the core and the steam generators follow.
A boiling water reactor operator has a lever a PWR does not: recirculation flow. Push more water through the core and you sweep out steam voids, which adds moderation, which raises power, without touching a rod. Power follows flow. The rods also enter from below, because the top of the vessel is full of steam lines going straight to the turbine. Forty-five are operating.
A CANDU operator refuels while the reactor runs, pushing fresh bundles through pressure tubes on load, and trims reactivity with liquid zone controllers rather than boron. That capability is forced by the fuel: natural uranium has so little excess reactivity that an eighteen-month batch cycle is not physically available.
None of these skills is portable. The taxonomy is the job description.
Where the same event goes differently
Take one fault: a stuck-open relief valve on the pressuriser. In a PWR, that is Three Mile Island — coolant leaves through a valve the indicator says is shut, the operators read rising pressuriser level as too much water and throttle back the injection that was saving the core. The instrument that lied and the mental model it fed are both PWR-specific.
A BWR does not have a pressuriser to fail that way at all. Its equivalent failures live in level control and recirculation. A CANDU’s live in the pressure tubes and the two shutdown systems. Every reactor accident in the catalogue here records the type it happened to, for exactly this reason: the failure modes are family traits.
Go and feel it
Reading about a positive void coefficient is not the same as watching one run away from you.
I built ReactorSim, a control-room simulator that runs all four of these families — PWR, BWR, CANDU and RBMK — with real physics and real procedures. You pick the plant, the scenario and where you are in the fuel cycle, then you operate it.
The scenarios are the ones that matter: loss of offsite power, small-break LOCA, steam generator tube rupture, a stuck rod on scram, the stuck-open PORV that defined Three Mile Island, a grid frequency disturbance that pulls load off your unit. There is a shift-turnover scenario where you inherit a plant that already has a protection channel bypassed and a clock running, and an unannounced-fault mode where nothing tells you what broke.
Run the same loss-of-feedwater in a PWR and then in a BWR and the difference stops being a table column. That is the argument for classification, made faster than any page here can make it.
Start with the reactor types index, then go break something.