Show Me Where That Is Written: Keep Your AI From Guessing

Never accept "that fits" from your AI: have it quote the line its answer rests on, and have it run a checker that shows the exact path of scenes. If it cannot show that, it should say it does not know instead of guessing.

Why does my AI say things fit when they do not?

An AI can sound just as sure when it is guessing as when it knows. Writing "that fits" costs it nothing.

Say you are building a small harbour mystery. You ask, "Can Hanne be the one who opens the sea gate?" The AI answers, "Yes, that fits her character and the timeline perfectly." You build the scene.

But the player can warn Hanne, and then she leaves on the last ferry. So the honest answer depends on the path: yes on one, impossible on the other. A quote alone cannot tell you that. Something has to walk both.

What counts as showing it?

• A line from the story bible, quoted, with its section. See what a story bible is.
• An earlier scene, named, with the line quoted.
• A rule, quoted from your world rules.
• Something you said, quoted from the chat.
• A path from the checker. The exact scenes that break a rule, or the number of states where every rule held.

These do not count: "it fits the tone", "as we established", "this should work" and "it feels right". None of them point at anything you can look up.

What does a real check look like?

Here is that harbour story as a model. The player warns Hanne or says nothing. Then comes the sea gate scene, where Hanne opens the gate. The rule: nobody acts where they are not.

We ran this invented story through TLC 2.19 on 2 November 2026. This is TLC's real output. We only left out its housekeeping lines and put each step on one line.

Error: Invariant NobodyActsWhereTheyAreNot is violated.
State 1: scene = "Start", happened = {}, hanneIs = "at the harbour"
State 2: scene = "Warn Hanne", happened = {}, hanneIs = "on the last ferry"
State 3: scene = "Sea gate", happened = {"Hanne opens the sea gate"}, hanneIs = "on the last ferry"

Read it from the top. On the path Start, Warn Hanne, Sea gate, Hanne opens the gate while she is on the last ferry. That path is the proof that the "yes" was wrong on this path.

The fix: if the player warned Hanne, Wim the lockkeeper opens the gate instead. We ran TLC again.

Model checking completed. No error has been found.
9 states generated, 7 distinct states found, 0 states left on queue.

TLC looked at all 7 states the story can reach, and the rule held on every path.

What if it cannot show it?

Then "I do not know" is the right answer, and a useful one. It marks the exact spot where a decision is yours. Decide it, write it into the story bible, and the gap is closed.

Can my AI still come up with new ideas?

Yes, and you want it to. It only has to say so. "This is new, and nothing in the story backs it up yet" is a fine answer. Once you say yes, the idea goes into the story bible and into the model, and from then on it gets checked like everything else.

Do I have to read any of this?

No. You never have to read the model or the checker's output, because your AI writes it, runs it and tells you the result in plain words. TLA+ is a language for modelling, and TLC is the tool that checks a model. If TLC cannot run on your computer, your AI can get a ready-made checker through our MCP server and run it in your project. Spielerando never looks at your story.

A pass means your rules held on every path in the model. It does not mean your story is good, and it says nothing about rules nobody wrote down.

Paste this to your AI

From now on, work from proof. This matters more than anything else I ask you. Never just tell me that something fits, is done or is true.

1. Read story-bible.md and the story so far. If you cannot see my files, ask me to paste them. If there is no story-bible.md, create one from what you find.
2. Write a TLA+ model of the story in a folder called story-check.
   - Facts, as state: who is alive, who knows what, who is where, who holds what, and what has happened.
   - Steps: every scene and choice, with what it needs and what it changes.
   - Rules, as invariants: every rule in story-bible.md, plus these: the dead do not speak. Nobody uses knowledge before the scene where they learned it. Nobody acts where they are not. No two scenes contradict.
3. Run the TLC model checker on the model so that it walks every path. If TLC cannot run here, use get_recipe with story-model-check through Spielerando and run that checker in this project. If you can run neither, tell me, and do not claim that anything was checked.
4. For each broken rule, show me the exact path of scenes the checker found, in plain words.
5. Whenever you say something about my story, show me what it rests on: the exact line, quoted, with where it is written, or the checker's result. Copy every quote exactly from the source. Do not write quotes from memory.
6. If you cannot show it, say "I do not know". Do not guess, and do not fill the gap with something that sounds right.
7. If you suggest something new, say "This is new". Add it to story-bible.md and to the model only after I say yes, then run the checker again.

Report only which rules held, how many states were checked and whether the check was cut off. Never tell me the story is good.

Start now: take the last three story decisions in this project, and show me where each one is written and what the checker says about it.

Or skip this and connect the MCP: your AI then finds all of this by itself.Connect the MCP

How to check it worked

• Say "show me where that is written". Pick something your AI just told you. It should give you a line you can find.
• Look up one quote yourself. An AI can get a quote wrong too. Search the file for a few of the words. If the line is not there, tell the AI.
• Play the path it gave you. Follow those exact scenes in the game. You should see the problem yourself.
• Ask about something the story never says. Ask what Hanne's mother is called. It should say it does not know.
• Ask a trick question. Ask why Hanne opened the gate on the path where she left. It should correct you and show the path.
• Ask for a new idea. It should tell you that the idea is new.