Deduction Strategy Guide
SherlockAI feels like magic when it lands a guess, but there is no trick involved: the game wins by asking mathematically good questions, and you can help it win faster by answering them well. This guide explains how the engine thinks and how to play each mode so rounds stay short and accurate.
How Information Gain Works
At the start of a round every candidate in the catalog is equally likely. Each answer you give is evidence: the engine lowers the probability of candidates that contradict your answer and raises the probability of candidates that agree with it. When it chooses the next question, it scores every unused clue by how cleanly it splits the remaining candidates. A clue that divides 300 surviving candidates into two groups of about 150 is extremely valuable; a clue that only affects three candidates is nearly useless at that point.
This is why the early questions feel broad ("Is this a real person?") and the later questions feel oddly specific. Broad clues are efficient when the candidate pool is huge, and narrow clues become efficient once only a handful of similar names remain.
Answering Borderline Clues
The five answer choices exist because real knowledge is fuzzy. Using them correctly is the single biggest thing you can do to improve the game's accuracy:
- Yes / No — use these when the answer is widely agreed upon. They give the engine a strong signal and can eliminate large groups of candidates outright.
- Probably / Probably not — use these when the clue is mostly true or mostly false for your target. They shift probabilities without hard-eliminating candidates, which protects the round when you are not fully certain.
- I don't know — use this freely. Skipping a clue costs one question, but a guessed answer that is wrong can mislead the engine far more than a skipped one.
The most common mistake is answering Yes or No to a clue you are unsure about. One confident wrong answer can eliminate your own target, and after that the engine can never recover because your answer told it the truth lies elsewhere.
Choosing Good Targets
The engine can only guess what is in its catalog, so target selection matters:
- Prefer famous, unambiguous names. A globally known footballer produces cleaner public consensus than a local celebrity, which means the catalog's traits match what most players would answer.
- Check the catalog first. The full character catalog and career catalog are published, so a quick search tells you whether your target exists before you invest ten minutes in a round.
- Commit to one answer. Switching targets midway contradicts your earlier answers and confuses the probabilities.
- Answer for the typical case. For careers, describe the ordinary version of the job, not an unusual specialist. For characters, describe the version most people know, not an obscure storyline.
Using Wrong Guesses To Your Advantage
A wrong guess is not a failure state. When you reject a guess, the engine removes that candidate and re-ranks the rest using everything it has already learned. In close categories, such as two athletes from the same sport and era, the first guess is often the second-most-likely name, and rejecting it points the engine straight at your target. Rounds with one or two rejected guesses still usually end faster than starting over.
Mode-Specific Advice
- Characters: think about how the general public sees the character, including origin, franchise, and role. Origin clues (comics, anime, video games, film, books) are doing most of the early work.
- Careers: answer based on daily duties, typical workplace, and common training routes. Clues about uniforms, shifts, and licences separate large groups of professions very efficiently.
- Sherlock's Profile: there is no hidden target, so accuracy means honesty. Answer as yourself and the generated case file will be noticeably sharper.
A Short Example
Suppose your target is a well-known video game mascot. The first answers (not a real person, not human, from a video game) collapse the 600-name catalog to a few dozen candidates within three or four questions. Clues about colour, abilities, and franchise then do the fine work. If the engine guesses the wrong mascot from the same franchise, one rejection usually settles the round. Total time: a couple of minutes, driven entirely by the quality of your answers.
Ready to test the theory? Start a round, or read the step-by-step rules first.