Roman numerals are the oldest notation most engineers still read weekly: film credits, regnal numbers, clock faces, chapter headings, Super Bowls. The system is seven symbols and two composition rules, and knowing which rule came first explains most of its oddities.
The seven symbols
I is 1, V is 5, X is 10, L is 50, C is 100, D is 500, M is 1000. That is the entire alphabet. There is no zero: the system counts things, and the Romans wrote the absence of things in words (nulla) rather than in numerals. There are no negatives either, and nothing larger than M.
Additive first, subtractive later
The original system was purely additive: symbols written largest-first and summed. Four was IIII, nine was VIIII, 1910 was MDCCCCX. The subtractive shorthand - a smaller symbol before a larger one subtracts, so IV is 4 and CM is 900 - existed in antiquity but only became the standard convention in the Middle Ages, long after Rome fell.
That history is why both spellings survive. Clock faces traditionally show IIII (visual balance with VIII across the dial is the usual explanation, and the Library of Congress keeps a delightful file on the question), monument inscriptions use additive dates, and yet every textbook teaches the subtractive canon. A well-behaved converter accepts the additive form, values it correctly, and shows the canonical spelling beside it - which is exactly what the Roman numerals converter does.
The rules the canon enforces
The subtractive canon is tighter than it looks. Only six pairs subtract: IV, IX, XL, XC, CD, CM. The pattern behind them: I subtracts only from V and X, X only from L and C, C only from D and M - each symbol reaches at most one order of magnitude up. That is why IL is not 49 (it is XLIX) and IM is not 999 (it is CMXCIX). Only a single symbol may be subtracted, so IIX is not 8 (VIII is). And the "five" symbols V, L, D never repeat and are never subtracted: two of them are simply the next "ten" symbol.
Each additive symbol repeats at most three times in the canon - III is 3, but 4 flips to IV. Stack the rules together and every value from 1 to 3999 has exactly one canonical spelling, which is what makes the notation parseable at all.
Why it stops at 3999
MMMCMXCIX is 3999, and the next value has no canonical plain-text form: there is no symbol for 5000. Roman practice used the vinculum, an overline that multiplied a numeral by 1000, so V-with-a-bar was 5000. Unicode carries combining overlines, but no plain-text convention for them ever standardized. A tool that computes rather than guesses states that limit instead of quietly inventing notation - the converter refuses 4000 and tells you why.
Reading 1994
The canonical construction is place by place, and reading it back the same way is the trick to fluency: MCMXCIV is M (1000) + CM (900) + XC (90) + IV (4). Each place is either a plain run or one subtractive pair, never both. Once you read numerals in places instead of symbols, MMXXVI stops being a puzzle and becomes 2026.