Build a Social-Deduction Puzzle Generator That Preserves Solvability

Add a puzzle generator that preserves solvability to your social-deduction game with AI. Integrate the system with hidden roles and evidence-based discussion.

Seele Editorial TeamUpdated November 18, 2026

Add a puzzle generator that preserves solvability to your social-deduction game with AI. Make state changes visible, preserve recovery, and test a repeatable scenario. Start with a small evidence model and a validator before scaling the number of roles, clues, or generated rounds.

Define puzzle constraints and evidence

Start with a compact case model: suspects, possible actions, time order, locations, and evidence that rules out or supports a theory. Write the intended solution as a set of constraints before generating clues. Decide whether a puzzle should have one valid conclusion or several defensible theories, then represent that target explicitly so the generator can test it.

A playable colorful sci-fi crew room with evidence fragments, task stations and a central discussion table scene frames the design problem.
Introduce the playable setting and challenge: colorful sci-fi crew room with evidence fragments, task stations and a central discussion table.

The official Among Us listing from Innersloth describes roles, tasks, suspicious behavior, and discussion. It supports the social-deduction context only; the puzzle generator here is original. Separate what each player sees from the hidden answer, and make sure the public evidence is sufficient for the intended deduction without exposing private role information.

Generate only valid states

Generate the hidden case first, then derive evidence from it. A validator should check that every clue is consistent with the case, every required role or object exists, and the final puzzle meets its intended uniqueness or ambiguity rule. Reject a seed that produces contradictory clues or no supported answer rather than patching it with a random hint after the fact.

A visual sequence shows the player action, system response and resulting option.
Show the cause-and-effect mechanism for generate only valid states.

For stronger verification, run a small solver against the generated evidence. Count the candidate solutions, then accept the puzzle only if that count fits the design: exactly one for a closed deduction, or a defined set of interpretations for an open discussion. Keep generation and validation separate so a failed seed can be discarded and logged with its reason.

Make clue deductions verifiable

Every conclusion should have a traceable evidence path. A clue can exclude one suspect, establish a timeline, or connect two observations, but the player should be able to point to the evidence that supports the inference. Avoid clues that depend on arbitrary wording or information no player could access in the current round.

The same game world shows a visible payoff caused by the earlier choice.
Illustrate the result of make clue deductions verifiable in the same game world.

Keep clue presentation consistent across generated cases. Use the same symbols for time, location, and role, and let players inspect evidence during discussion. If the puzzle allows multiple theories, show which facts remain uncertain and let the group compare competing explanations without one hidden rule invalidating reasonable reasoning.

Test seeds and edge cases

Run a batch of seeds through the validator and inspect accepted and rejected cases. Record solution count, clue coverage, contradictory evidence, and whether a player can reconstruct the intended answer from available clues. Test cases with missing evidence, repeated suspects, tied timelines, and role-specific information so edge conditions do not produce an impossible round.

For related rule and evidence design, see a guide to original puzzle rules with multiple valid solutions and a guide to environmental puzzle rules. Pass when every accepted seed has the intended solution count and players can support conclusions with available evidence.

FAQ

How should a social-deduction puzzle generator work?

Generate a hidden case from explicit constraints, derive evidence from that case, and validate the resulting puzzle before presenting it. Decide whether the round should have one answer or a defined set of defensible interpretations. Reject seeds with contradictions, missing required information, or a solution count that does not match the design.

How can generated puzzles remain solvable?

Run a validator or small solver over the public evidence and count the candidate solutions. Accept a puzzle only when the count matches the intended format, such as exactly one conclusion or a bounded set of interpretations. Keep generation separate from validation so failed seeds can be discarded with a recorded reason.

How should evidence support deduction?

Make each clue traceable to the hidden case and useful for ruling in or ruling out a theory. Establish facts such as time, location, or action through consistent evidence, and ensure players can access the information needed for the intended conclusion. Avoid arbitrary clues or hidden rules that invalidate reasonable deductions.

What should you test in a puzzle generator?

Validate many seeds and inspect solution count, clue coverage, contradictions, and whether players can reconstruct the intended answer. Include edge cases such as missing evidence, repeated suspects, tied timelines, and private role information. Confirm that rejected seeds are logged and that every accepted seed meets the chosen uniqueness or ambiguity rule.

SEELE AI

Turn this guide into a playable build.

Start with a focused brief, then iterate in the Seele AI Workspace.

Start creating