Following one turn from your input to the narration

Say a Player types "I draw my sword and step toward the guard" and presses send. A second or two later, prose starts arriving. In between, the platform does a fair amount of work, and most of it has nothing to do with writing. This article follows that single turn from start to finish, in the order things actually happen.

The short version: several plain checks sit on both sides of the story model, and it never changes your game directly.

Step one: your input is checked before anything else reads it

The first stage looks at what you wrote against the campaign's safety settings, with a small window of recent turns for context. If the input is blocked, you see a block message and the turn goes no further. Nothing reaches the rules, the story model or the saved campaign.

Only input that passes moves on.

Step two: your words become a structured intent

Free text is hard to check, so a model reads your input and produces an intent: what action you are attempting, who is doing it, who or what it targets, and whether it is mechanical (an attack, a lockpick) or purely narrative (looking around, chatting).

That intent can mention items. If a campaign tracks inventory, the intent is built with your actual inventory in view, and then a separate step checks every item the model claims you used against what you really carry. A claim that does not match is dropped, not trusted. Claims about picking up or losing items get the same treatment, and anything unmatched is discarded rather than made true.

A model can propose. The platform decides whether the proposal holds.

Step three: the rules resolve the action

If the action is mechanical, the campaign's ruleset evaluates it. The ruleset is a separate, sandboxed piece of code, and the story model has no part in this. The outcome is plain: it worked or it did not, and which values change.

The ruleset can reject the action, for example because the verb is not one it understands. A rejection is an expected result, not a crash: you get a clear "not allowed" response, nothing is narrated, and nothing is saved. Purely narrative actions skip the ruleset entirely, so looking around is never refused by a combat rule.

Step four: the proposed changes are held, not applied

Once the outcome is known, the platform records the pending changes as a draft before the narrator is called. The story model can take a while, and the platform does not hold a database transaction open while it waits. The draft marks the turn as in progress, so a failure can be settled honestly instead of left half-finished.

Step five: the narrator writes, with the facts handed to it

Now the story model is called. It receives the mechanical outcome, the relevant canon, character information and recent history, with your input and other dynamic content wrapped in clearly marked tags so that it reads as material to narrate and not as instructions to follow.

The prose is collected in full before you see any of it. Nothing is streamed to your screen while the checks below run.

Step six: the finished prose is checked twice

The first check is an output safety filter on the complete text. If it fails, the platform asks for one new attempt with an added safety instruction. If that attempt also fails, you get a block message instead of the text.

The second is a consistency check, and it is narrower than the name suggests, so it is worth being precise. It does not grade style, tone or general plausibility. It looks for contradictions against specific facts the platform already knows for this turn. For example, a secret held by a character that the narration must not have them reveal, or, when inventory tracking is on, a load-bearing item the prose has you using that you do not own.

When a contradiction is found, the platform regenerates the narration with the contradicted facts called out, up to three attempts. A rewritten version has to clear the safety filter again before it is allowed through. If all attempts still contradict the facts, you are told the turn could not be narrated consistently, rather than being given text the platform does not trust.

One honest limit: if the consistency checker itself is unavailable, the turn goes ahead unchecked and the gap is logged. It is a continuity aid on top of the other checks, not a gate that blocks every turn when it is down. When there is nothing to check, the call is skipped.

Step seven: the story arrives, and only then does it become canon

Only now is the prose sent to you. Then the pending changes are applied to your character and the world, together with the journal entry and the finished text, in a single atomic step. Either all of it lands or none of it does. If the commit fails, the draft is marked unusable and you are told so.

What this adds up to

Input can be blocked, claims dropped, actions rejected, prose rewritten or refused. None of this makes narration perfect, but a bad turn tends to fail visibly or get fixed quietly, rather than silently rewriting your campaign.

If you want to see it from the Player side, start a campaign from Discover and try an action the ruleset might refuse. The response will tell you which path your turn took.