"Evil" means something specific here
On most sites, evil is simply the label above expert — a mood rather than a measurement. We use it for a precise thing: a puzzle that cannot be finished with any of the patterns on the expert page, but can be finished with an almost locked set or a swordfish. Every puzzle served here has been individually checked against that definition.
That check mattered more than we expected. The raw collection these came from turned out to be badly mixed: a quarter of it genuinely met the bar, while a good deal was easier than our expert puzzles and a large slice was harder than the label suggested. Everything that didn't belong has been moved or dropped, which is why this bank is smaller than it started — and why the word means something.
You never need to guess on this page — if you're stuck, a findable pattern exists, however long the finding takes.
Swordfish
If you know the X-wing, you already understand this — it's the same idea one size up.
An X-wing is a digit confined to the same two columns across two rows. A swordfish is a digit confined to the same three columns across three rows. The rows don't each need all three columns; two of the three is fine, as long as between them the three rows only ever use those three columns.
Three rows will consume the digit three times, and they can only do it inside those three columns. So the columns are fully accounted for, and the digit can be struck from them everywhere else. It works with rows and columns swapped, and the awkward part is purely visual: nine potential squares, rarely more than six actually present, spread across the grid.
Almost locked sets
This is the technique that does most of the work at this level, and it's less known than it deserves to be.
A locked set is n squares in a unit holding exactly n candidates between them — three squares holding only 2, 5 and 8, for instance. They're sealed: those three digits go in those three squares.
An almost locked set is one candidate short of being sealed: n squares holding n+1 candidates. It isn't locked yet, but it is one elimination away. Remove any single one of its candidates and the rest snap into place.
Now take two such sets that share two digits, X and Z, where every X in the first set sees every X in the second. Only one of the sets can supply the X. The other is therefore starved of an option, locks, and is forced to supply the Z. You don't know which — but either way, a Z lands inside one of the two sets. Any square that can see all the Z candidates in both sets therefore cannot be a Z itself.
Finding one without losing an hour
Start with bivalue squares — a square with exactly two candidates is the smallest possible almost locked set, and pairs of them are where most usable examples come from. From there, look at units with three or four empty squares between them holding one more digit than they have squares.
Then check for a shared digit whose positions in the two sets all see each other. That mutual-exclusion link is the rare part, and it's what makes the whole thing work.
How long these take
Considerably longer than expert, and the time is spent searching rather than deducing. An evil grid will typically open normally, stall completely, and then need a deliberate, systematic hunt. Expect stretches where you make no progress at all.
That's the intended experience, and it's worth saying that "no progress" here is not the same as "no path" — the path is guaranteed to exist.
If you want to meet either of these on demand rather than waiting for one to turn up, the technique trainer will find you a swordfish position, or an almost locked set, with everything easier already played out.
If it's still too gentle
Extreme is a different proposition entirely: puzzles where every technique on this site, including everything above, runs out. Those need chains, or a decision to test something and see what breaks.