Before the Spell Becomes Canon: Validating Generative Facts

From Magical Mysteries to Mathematical Meaning · Article 4 of 7

Once a mystery has a persistent truth network, it becomes tempting to think the hard problem is solved. The model now has somewhere to store facts. Characters, events, observations, beliefs, lies, and clues can all be represented outside the conversation. Surely the model can simply continue generating into that structure.

That turns out to be dangerous. A structured state does not govern itself. If every plausible detail produced by the model is immediately inserted into canonical state, the system has only moved uncontrolled improvisation from prose into data. The representation is more explicit, but the project can still drift.

The crucial next step is to separate proposal from commitment. A newly generated fact should not become true merely because it is linguistically plausible. It should first exist as a candidate mutation to the world. The system can then ask what that mutation would do to everything already committed.

Consider a simple example. The player enters a previously unexplored bedroom and asks what is on the desk. The model proposes an unopened telegram addressed to one of the suspects. In ordinary interactive fiction, the telegram would simply appear. But the existence of that telegram may imply a communication channel, a sender, a delivery time, knowledge available to the recipient, a reason it remained unopened, and perhaps information that changes motive or opportunity. A locally attractive detail can modify the case globally.

So the system pauses before narration becomes authority. It identifies the affected objects and relations. It checks whether the sender could have sent the telegram when the world says it was sent. It checks whether the addressee’s prior statements remain coherent with receiving it. It asks whether the message reveals a fact the suspect should not yet know. It examines whether the telegram duplicates another clue, undermines an alibi too early, or creates a second solution path that the rest of the case does not support.

Only after those checks does the candidate become a committed fact. The narrative can then safely tell the player that the telegram exists.

The candidate state can therefore have its own lifecycle. A detail may be proposed while it is being considered, validated once coherence checks pass, and committed only when the project accepts it as authoritative. If a later revision replaces it, the old fact can be superseded rather than silently erased. This makes the history of the world model inspectable without confusing superseded states with current truth.

This is a transactional view of generative world-building. The useful sequence is roughly:

read current state → propose mutation → identify dependencies → validate invariants → revise or reject if necessary → commit → propagate consequences → materialize prose.

The order matters. If prose is generated before the state transition is accepted, the system may narrate itself into a contradiction. Once the player has seen the telegram, quietly deleting it from the canonical data is not a clean correction; the player’s experienced history now contains a fact the world no longer recognizes.

This distinction is easy to miss because language models are so good at producing text that feels authoritative. The protocol has to resist that surface effect. A fluent sentence is still only a proposal until the state machinery accepts what it says.

Validation is also broader than contradiction detection. A candidate fact can be logically compatible with the existing world and still be bad for the mystery. Suppose the system proposes that a witness saw the murderer’s face clearly while leaving the scene. Nothing in the timeline may contradict that observation. The witness may have been in the right location, looking in the right direction, under adequate lighting. Ontologically and temporally, the fact is fine.

Epistemically, however, it may destroy the case. If the witness is reliable and immediately available, the observation collapses the space of hypotheses before the player has done any meaningful investigative work.

The validator therefore needs to consider several kinds of coherence. There is ordinary world coherence: can these facts all be true together? There is temporal coherence: can the represented events occur in this order? There is causal coherence: do effects have compatible causes? There is epistemic coherence: can each character know, believe, infer, or lie about what the state says they do? And there is evidentiary coherence: does the distribution of available information preserve the intended path toward justified solution rather than making the case impossible or trivial?

Some of these checks can be formal. If the schema represents time intervals and exclusive locations, an impossible overlap can be detected mechanically. If an object has a single owner at a time, contradictory possession states can be flagged. If knowledge acquisition is represented explicitly, the system can identify a claim that a character knows something before any represented path could have supplied it. If the case uses dependency edges, a changed fact can trigger an impact traversal to identify clues, scenes, or hypotheses that require review.

Other checks remain semantic. Whether two differently phrased statements actually contradict each other may require interpretation. Whether a new motive fits a character’s established behavior is not reducible to graph topology. Whether a clue is too revealing depends partly on the whole rhetorical and epistemic shape of the case. The model remains essential because the state does not replace semantic inference; it gives semantic inference a bounded, inspectable substrate.

This hybrid division of labor is one of the central strengths of the method. The model proposes and interprets. The formal state preserves identity, topology, and authority. Deterministic operations handle whatever the representation makes exact. The protocol decides when a change is allowed to become canonical.

The same cycle also makes revision tractable. Suppose late in development I decide that the culprit used a concealed stair rather than the warded conservatory. In a prose-first mystery, that change may require manually rereading every scene and hoping to notice all consequences. In a structured system, the stair-route fact can be treated as a mutation to canonical state. The system can traverse relations to find footprints, witness sightlines, timing assumptions, map descriptions, alibis, and clue interpretations that depend on the old route.

Not every dependent artifact needs complete regeneration. Some may be unaffected. Some may require factual correction. Some may need a new clue. Others may require recomposition because their function in the evidentiary trajectory has changed. The graph gives revision a scope.

That is a different notion of generative flexibility than simply allowing the model to rewrite anything at any time. The flexibility comes from being able to change a governed representation and then recompute its consequences.

The player can still surprise the system. That is important. A role-playing mystery becomes dull if every interaction has been pre-scripted. The player may search an unexpected location, question a minor character, test an unusual hypothesis, or ask whether a physical object has a feature the designer never anticipated.

The protocol does not need to prohibit those moves. It gives the model a way to respond without sacrificing the integrity of the case.

If the player searches beneath the greenhouse floorboards, the system can generate candidate findings consistent with the current world. It might decide there is nothing relevant there. It might create a mundane historical detail that enriches the setting. Or it might notice that an existing world fact naturally implies a trace that could plausibly be discovered there. In each case, the proposed addition is checked before it is promoted.

The important boundary is between semantic invention and world commitment. That boundary generalizes well beyond mystery games. Any long-running LLM project accumulates decisions that should not be casually rewritten by later local generation. A research workflow has claims and sources. A software project has interfaces and invariants. A curriculum has learning objectives, prerequisite relationships, examples, and assessments. A writing project has established arguments and terminology. In each case, the model benefits from being able to propose changes while a protocol protects the larger state.

The mystery application makes this easy to see because contradictions become dramatically visible. A suspect cannot be in two rooms at once. A witness cannot know what they never learned. A clue cannot sensibly derive from an event that never happened. But the deeper principle is about project-scale generation: new local output should be evaluated against durable global structure before it becomes authoritative.

That principle also changes the status of prose. Prose is no longer the primary carrier of project truth. It is a derivative artifact produced from accepted state. If a paragraph, room description, witness statement, or instructional explanation contains a new substantive commitment, the state should either already support it or be updated through the mutation protocol that makes it true.

Once this architecture was in place, the magical-mystery system stopped looking like a special-purpose trick for wizard-detective fiction. It had a target terminal state. It modeled plausible initial states. It constructed an evidence-bearing path backward. It represented the relevant domain as persistent formal-semantic state. It governed local mutations against global constraints. And it materialized an experience that a human traversed forward.

That combination suggested a more general description. The mystery was an epistemic machine.