You are shown a colour word printed in an ink colour that may or may not match it — the word RED rendered in blue. Your task is to report the ink colour and ignore the word. Reading is so heavily practised that the word's meaning arrives whether you want it or not, and when meaning and ink disagree the two answers compete for one response. Stroop's 1935 experiments established the interference with colour names in conflicting ink, and it has been replicated often enough to be a standard demonstration of automatic processing.
Exactly 20 trials are generated — 10 congruent (word and ink agree) and 10 incongruent (they disagree) — then shuffled, so you cannot predict what is coming. The palette is red, blue, green and orange, and the four answer buttons always sit in that fixed order. Timing starts as the stimulus is drawn and stops on your click; each response is followed by roughly 450 ms of feedback. The result reports two averages, both using correct answers only: the congruent mean and the incongruent mean. The headline is their difference, interference = incongruent mean − congruent mean, in milliseconds — how much extra time the mismatched words cost you.
Suppose you answered all 20 trials correctly. Your congruent times were 612, 598, 641, 605, 630, 588, 654, 622, 636 and 601 ms, which sum to 6187, so the congruent mean is 618.7 ms. Your incongruent times were 742, 705, 781, 733, 768, 719, 796, 751, 749 and 793 ms, summing to 7537, for a mean of 753.7 ms. Interference = 753.7 − 618.7 = 135 ms, above the 100 ms mark the tool treats as a noticeable pull.
Now suppose the 796 ms trial had actually been an error. Errors are dropped before the averages are taken, so the incongruent mean would become 6741 ÷ 9 = 749.0 ms and the interference would fall to about 130 ms. One wrong answer moved the score by 5 ms — the speed–accuracy trade-off in miniature, since answering faster by guessing produces errors, and errors shrink the very mean the score rests on. With only ten trials per condition, a 100 ms outlier shifts that mean by just 10 ms.
Neither — it is a size typical of an adult on a mismatched-colour task, and the page has no norms. What is informative is your own pattern across repeated runs under the same conditions.
An error means you responded to the wrong thing, and averaging errors in would mix speed with accuracy. Accuracy is reported separately so you can judge whether the timings are trustworthy.
The interference happens before you respond. You had to resolve a conflict the matching trials never created, and that resolution is what the extra milliseconds represent.
That is the standard interpretation: for a practised reader the printed word is processed automatically, so the task cannot be done by filtering it out — you can only override the competing response.
You start trading correctness for speed. A fast run with several errors shows a shorter incongruent mean and lower accuracy at the same time, so read the two figures together.