Source notes
Whereabouts
How this project works, the decisions behind it and its known limits. Ask quotes and cites these passages when it answers.
Notes reviewed 20 September 2026
Explore the projectProduct
Product
A daily mystery. A theory is not enough: prove it. Whereabouts is a daily deduction game built around the satisfaction of changing your mind and then proving what happened. The goal is a fair, human, memorable case that makes the player want another one tomorrow.
The problem it solves
A generated mystery can sound convincing while containing an impossible answer, a hidden assumption, an obvious giveaway or simply a dull pile of documents. Fairness and entertainment are separate requirements. A correct answer alone also does not show whether the player understood the evidence.
The experience
The player reads a case, inspects people and artifacts, develops a theory, and submits both an answer and supporting proof. Authoring and player-facing state are separate. Behind the product, Generator Studio designs and critiques candidate cases before they can be admitted for publication.
Engineering
Backend and architecture
The web player uses React and TypeScript, with a separate SwiftUI client. Supabase Postgres and Edge Functions hold authoritative game state and publication boundaries. The deduction engine operates over a finite authored world and supported rules, deriving conclusions from selected evidence. Generator Studio is an orchestration layer: concept exploration, an independent experience director, case design, evidence authoring, deterministic fairness gates, a sequential blind player, independent reviewers and a ship decision.
Technology by responsibility
Client: React, TypeScript, SwiftUI. Backend: Supabase, PostgreSQL, Edge Functions. Generation: Structured model tasks, Deterministic logic engine, Review and feedback orchestration.
Data, trust and failure behavior
The generator pipeline returns private candidates and receipts; it does not itself grant publication. Immutable revisions, admission rules and the authoritative judge form a separate release boundary. The finite logic engine only executes supported rule types and derives from the evidence actually selected. A true blind transcript must not cite unopened evidence. The player and the generator are developed on separate branches, so a generator change is not automatically in the game people play.
Theory plus proof
The logic model represents selectable artifacts, propositions and deduction steps. It can distinguish an entailed conclusion from incomplete or contradictory evidence.
Design before writing
Concept Forge explores alternatives and an Experience Director can reject all of them before the expensive case-writing stages.
A genuinely sequential playtest
The blind-player protocol reveals one chosen artifact at a time rather than giving a reviewer the complete solution and calling the result a playtest.
Repair the cause, not the score
The feedback loop distinguishes a writing defect from an unsupported difficulty tier or a systemic kernel/pipeline problem. It can redesign, retier without rewriting, or stop rather than lowering the quality gate.
Decisions
Separating a fair puzzle from a fun one
Deterministic checks establish properties of the declared logic; independent player and editorial review evaluates curiosity, voice, reinterpretation and whether the case earns another session. Trade-off: A formal proof over authored propositions does not automatically prove that the prose communicates them fairly. The symbolic rules and the human reading experience are therefore checked separately.
Learning from rejection without endless rerolls
The feedback loop preserves the selected concept, carries review findings and general lessons into bounded repairs, and records costs and stage outcomes. Systemic failures stop the writer loop. Trade-off: A more deliberate process costs more than one-shot generation and can accept no cases. That is preferable to silently publishing weak content or treating a completed model call as a quality result.
Limits
Known limits
The architecture is substantial, but it does not prove that the generator consistently produces enjoyable cases, and no acceptance rate or commercial-quality guarantee is claimed. Some older documents and cached pages still use the product's historical name.
Questions Ask answers from here
Is a logically valid case automatically a good game?
No. The system deliberately separates truth and fairness checks from player and editorial quality. A valid but dull case should still be rejected.
What does the blind player see?
The current protocol starts with the opening and artifact labels, then reveals only the artifact the player chose. The resulting sequence is evidence for the ship decision.
What happens when a case fails review?
The feedback loop diagnoses the owning layer. Writing can be revised, difficulty can be corrected, and systemic problems can stop generation rather than prompting another blind reroll.