Inside the Engine: How SherlockAI Deduces Your Character
When SherlockAI names your character after a dozen questions, it feels a little like mind reading. It is not. Under the interface is a small, transparent probability model with no improvisation and no shortcuts — just a well-organized catalog and a disciplined way of spending each question. This article explains how it works, in plain language.
Everything starts as a list of candidates
At the beginning of a character round, the engine does not know anything except the shape of its catalog: 600 playable names, from athletes and musicians to comic-book heroes and figures from mythology. Each of those 600 entries starts the round as a candidate with an equal slice of probability. Your secret character is, from the engine's point of view, hiding somewhere in that list.
What makes the list searchable is that every entry is tagged by hand against the same set of traits. Is this a real person? Are they alive? Are they associated with film, with comics, with a sport? Do they fly? Do they use magic? The current character catalog tracks more than fifty such attributes for every entry, and the career catalog does the same for professions with workplace traits like shift work, licensing, and public-facing duties. These tags are the raw material of every question you will be asked.
Your answers are evidence
Each time you answer a clue, the engine walks through the entire catalog and re-scores every candidate. A character that clearly matches your answer gains probability; one that clearly contradicts it loses probability; and the strength of the update depends on which of the five answer choices you used.
This is why the five choices exist. A firm Yes or No is strong evidence and moves probabilities a lot. Probably and Probably not are softer nudges for clues you are unsure about — they guide the engine without letting a shaky answer eliminate your own target. I don't know skips the clue entirely, trading one turn for safety. After every answer, the probabilities across the catalog are renormalized, so the engine always has a live ranking of who your character is most likely to be.
Choosing the next question: information gain
Here is where the engine earns its name. With hundreds of candidates still alive and dozens of unused clues, which question should come next?
The engine scores every remaining clue by its expected information gain — a measure, borrowed from information theory, of how much the answer is expected to shrink the uncertainty. Intuitively, a clue scores well when it splits the surviving candidates into two substantial, roughly balanced groups. A clue that almost everyone matches, or almost nobody matches, scores poorly because the answer is nearly a foregone conclusion.
You can watch this logic from the player's seat. Rounds open with broad dividers — real or fictional, human or not, which part of the world — because those clues bisect a 600-entry catalog efficiently. Late in a round, when only a handful of similar names remain, the questions turn oddly specific, because only a fine-grained clue can still separate, say, two footballers of the same era. The questioning style that feels natural to a good human player is exactly the style that information gain rewards.
Guesses, rejections, and second chances
When one candidate's probability pulls far enough ahead, the engine stops asking and offers a guess. If you reject it, that candidate is removed and the round continues with the probabilities already computed — which is why a rejected first guess often converts into a correct second guess almost immediately. The runner-up was usually right behind.
If the catalog genuinely does not contain your answer, the round ends by inviting you to submit the missing name. Those suggestions are reviewed and, when accepted, tagged against the same trait set as every existing entry, so the catalog grows without breaking its own structure.
Why authored data instead of improvisation
A fair question is why the game does not simply invent questions on the fly. The answer is reliability. An improvised question might be ambiguous, factually wrong for half the catalog, or impossible to score against the trait data. Every clue in SherlockAI was written against the catalog, validated by automated checks, and reviewed before it shipped. The trade-off is real — the question pool is finite and curated — but it means the engine never asks you something it cannot interpret the answer to.
The same philosophy explains where the computation happens: in your browser. The catalog and the engine run locally on your device, so a round works without an account, answers are not sent off to be interpreted by a remote service, and your statistics stay in your own local storage. The design goal was a deduction game that behaves like a well-made physical game: everything you need is in the box.
Try it with new eyes
Next round, pay attention to the order of the questions. The first few will feel broad, the last few almost eerily specific, and a rejected guess will be followed by a visibly smarter second one. That is the whole trick — candidates, evidence, and information gain, repeated twenty times. Start a round, or read the strategy guide to learn how to make the engine's job as easy as possible.