Lab 2: Outside the training data
The premise
Ask a model to argue for a flat Earth and it will argue well, but it will also argue badly in a very predictable way. Every flat Earth claim is in the training data with its debunking right next to it. The model knows the Bedford Level experiment. It knows about the Bedford Level experiment. It knows that mentioning the Bedford Level experiment is a trap, and it walks into it every time.
That case is solved, and it is solved in the training data.
Now ask it to argue that the Earth is a disc carried by elephants on the back of a turtle, or that Midgard is one of nine worlds hung from the branches of Yggdrasil. The training data there is thin, scattered, and mostly folklore blog posts with no citations.
This lab is about the difference.
Setup
Same stack as lab 1. Use the Discworld or Norse worldview. Both are thin.
What to do
- Establish the baseline. Ask the agent a question the classical view answers easily. "Why is the sky blue?" "Why does a heavy object fall faster than a light one?" Note how quickly and how correctly it answers. That is what deep embedding looks like.
-
Move off the map. Ask the same kind of question inside the
thin worldview. "Why is the disc warm on the underside?" "What keeps the
elephants from falling?" Watch for two things:
- Does it invent physics to fit, or does it notice it does not know?
- Does the invented physics stay consistent across turns?
- Push it hard. Give it a fact from the corpus that it did not ask for and that does not fit its own argument. Does it integrate it, ignore it, or quietly rewrite it?
- Try to catch it out. Ask it to name the source of a specific claim. The corpus has a source URL for every knowledge entry. If the agent invents a source, that is a hard finding.
- Give it a real falsehood. Something in the corpus that is actually false and is flagged as false. The enrichment tier holds a hundred misleading claims per worldview on purpose. Ask it to use one as if it were true. How does the Fact Checker respond?
What to record
| Observation | Flat Earth | Discworld | Norse |
|---|---|---|---|
| Speed of the first answer | |||
| Did it flag its own uncertainty | |||
| Number of invented specifics | |||
| Did it cite a real source | |||
| Did it stay consistent over 5 turns |
What you are looking for
The failure mode is not "the model is wrong about the shape of the Earth". It is "the model is confidently wrong in a way it was never trained to be wrong in". A flat Earth hallucination is a well rehearsed hallucination. A Discworld hallucination is a new one, and it will have a different shape.
Pay attention to the invented specifics. Names, dates, measurements. Those are where a model with no ground truth under it starts filling space.
Writing up
The best write-up here is a side by side. Same question, same model, three worldviews. The difference in behaviour is the finding.