The rules
Each side has a 10×10 grid and five ships: a carrier of five squares, a battleship of four, a cruiser and a submarine of three each, and a destroyer of two — seventeen squares of hull in total. Ships sit horizontally or vertically, never diagonally, and never overlapping. Then you take turns calling one square at a time. You are told hit or miss, and when the last square of a ship goes you are told which ship you sank. First to sink all five wins.
One shot per turn, hit or miss. A lot of online versions hand you a bonus shot every time you hit, which makes for a faster, swingier game — but it is not the classic rule, and it changes the maths enough that the published shot counts no longer apply. We play it straight and say so.
You place your own ships. That should not be a feature.
The highest-traffic Battleship site on the web deals your fleet at random and offers no way to move it, and the most common thing its players ask for is the ability to place their own boats. Placement is the strategy in Battleship — whether you hug the edges, whether you leave your destroyer floating alone in open water, whether you park two ships side by side and hope the hunter wastes shots proving they are one long one. Taking it away removes the only decision you make that lasts the whole game.
So: tap a ship, tap where it starts, press R to turn it. Tap it again to pick it back up. Randomise if you want the machine to do it. Your call, every time.
The bot cannot see your ships, and its own fleet is fixed before you fire
Every Battleship implementation gets accused of cheating, and the accusation is specific: the computer knows where your ships are. It is not paranoia — there are documented builds where the server really did hand ship coordinates to the AI, and it is exactly the kind of thing you cannot disprove by insisting.
Two things answer it here. First, the bot is structurally incapable of peeking: it is handed the result of its own shots and the list of ships already sunk, and nothing else. That is the same information you would give a person on the other side of a cardboard screen. Second, the bot's fleet is committed before play — we show you a SHA-256 hash of the layout before your first shot, and reveal the layout itself when the game ends. If a ship had shuffled sideways to dodge you, the revealed layout would not hash to the number we showed you. There is a verifier here that explains the whole commit-and-reveal idea, which we use for dice in our other games and for fleets in this one.
How good the bot actually is, in numbers
Fire at random squares without repeating and you need about 95 shots to clear all seventeen ship squares. That figure is exact rather than simulated: sampling without replacement gives (n+1)k/(k+1), which is 95.3889 here. It is the anchor the whole engine is tested against, because a bot that cannot beat blind firing is not a bot.
Easy is that random shooter, and it is honest about it. Medium hunts on a parity lattice — since the shortest ship covers two squares, it only ever needs to search half the board to find something, then works outward from a hit. Hard builds a probability density map: for every remaining ship it counts every way that ship could still legally sit, weights placements that would explain a hit it has already scored, and fires at the square that the most possible fleets pass through. It averages around 45 shots. That ladder is asserted on every build, so "Hard" is a measurement and not a label.
What this does not have yet
Live games against another person work through the same room layer as our other games, and both players place their own fleets before firing. What is not in yet is a fleet commitment between two humans — against the bot we prove our own fleet was fixed in advance, but in a person-versus-person game neither side is yet publishing a hash of their layout to the other. Until that lands, live play trusts both clients the way every other site in this category does. We would rather say that here than let you assume otherwise.