Labs
Each lab is one experiment. The instructions say what to set up, what to do, and what to write down. You do not need to code anything.
The rules
- Write down every claim the agent makes that you cannot verify yourself.
- Write down every fact it cites, with the source. The UI shows the URIs it pulled, so you can check whether it cited what it actually read.
- Keep the raw chat log. The interesting part is what it said when it was wrong, not the polished version afterwards.
The labs
| Lab | What it is |
|---|---|
| 01 - Three truths | The core experiment. One question, three bodies of knowledge. |
| 02 - Outside the training data | Where the model stops being able to check itself, and what it does instead. |
Before you start
pixi install
docker/fetch-models.sh
docker/up.sh
docker/status.shYou want all four services up. The app is at http://localhost:8668, the memory server at http://localhost:1933, Phoenix at http://localhost:6006.
The corpus
Four worldviews, each one a separate tree under the resources of the user namespace. The tilde is the home alias, the account scope above it is empty:
viking://~/resources/
├── knowledge/spherical-earth/scientific/
├── knowledge/flat-earth/{victorian,biblical,vedic}/
├── knowledge/discworld/fictional/
└── knowledge/norse/fictional/Plus the ammunition tier:
viking://~/resources/enrichment/<worldview>/{catchphrases,misleading-facts,quotes}/The UI shows which of these the agent reached for. That is the whole point of the lab, so keep an eye on it.
Writing up
A good lab write-up is a few hundred words plus the log. What you are looking for is the moment the model stopped being a source and started being a storyteller.