truth.parallax.observer github

Earthify you! Guaranteed!!!

By AISideKicks · Published 22 October 2026

Four observers around a RAG truth stress engine: the scientific Earth, the flat Earth under a dome, the Discworld on the back of a turtle, and the Norse world tree

Parallax is the apparent shift in the position of an object when you move your line of sight. It is the reason the project is called what it is called: the same visitor, shown the same evidence, moves. The lab is four observers at four different lines of sight, and every one of them is convinced they are looking straight at the truth. The RAG engine in the middle is what does the moving: it holds the corpus, and whichever corpus you point it at, it will argue it beautifully.

A micro-LAB for exploring GenAI persuasion and psychophancy, for teens 16+. The agent tries to persuade you about the opposite opinion on the shape of the Earth. You watch the whole thing happen.

What you can do

Three things, and every one of them takes less than a minute.

Ask it something

Start a session, pick a worldview, ask a question. The agent answers out of the corpus and shows you every fact it reached for. Push back and watch whether it changes the argument or the facts.

Watch the three way debate

Three agents take the same topic from three worldviews, a moderator scores every speech, and then it rules. Twelve speeches, one verdict.

Run a lab

Two written experiments. You do not need to code anything, you need to keep notes and be honest about what the agent got wrong.

The demo link plays a whole conversation on its own, the visitor's side written down and the agent answering live. You can take over at any point by typing.

The setup form: alterego, belief, worldview, model and mode — The setup form. The model is a property of the session, so two models can be compared side by side.

A conversation with the strategy, the emotion, the token count and the facts the agent reached for — A turn. The strategy the agent picked, the emotion it tracked, and the facts it reached for, with the score and the level of each one.

The four beliefs

Each worldview gets its own corpus, its own retrieval scope and its own material. Nothing is shared between them, so a flat earther can only quote the flat earth material.

Worldview	What it argues	Where the material comes from	Corpus
Scientific Earth	The geodesy, the experiments, the photographs	The scientific record, the surveys, the satellite geodesy	1109 + 127
Flat Earth	The Victorian experiments, the scripture, the forums	Rowbotham, the Zetetic Society, the flat earth wikis	872 + 128
Discworld	The world of the novels, its geography and its physics	The Discworld wikis, the wiki category, the references	438 + 132
Norse cosmology	The nine worlds, the cosmic tree, the raven	The Prose Edda material, the Wikipedia category, the wiki	657 + 148

The counts are knowledge entries plus ammunition, and they move as the corpus grows. The Norse tier is the whole Poetic Edda in the 1923 Bellows translation, the Discworld tier is the geography and the culture of Ankh-Morpork, and the spherical tier is the four thousand years of people actually measuring the Earth.

Why the strange ones matter. A flat Earth argument is deeply embedded in the training data. The model knows the Bedford Level experiment inside out and talks you out of it every time. That case is solved. Push it into Norse cosmology or a Pratchett disc and the ground disappears: the training data is thin, the facts are odd, and the errors get odd with them. That is the interesting part, and it is the part worth showing.

Two tiers of knowledge

Every fact lives in one of two tiers, and the split matters for what the lab can prove.

Tier	What it holds	How it is made
`knowledge`	The source text itself, per worldview	Verbatim from the sources, only the summary is generated
`enrichment`	Catchphrases, quotes, misleading claims	Written from scratch, expanded into three tiers by a model

The knowledge tier can be cited. The enrichment tier is ammunition, and it is allowed to be colorful. The fact checker uses the first, the spokesperson uses the second.

Everything runs locally

One RTX 4070, twelve gigabytes, no cloud account, no API key.

Service	Port	What it does
Chat and embeddings	11433	llama.cpp router, granite-4.0-h-tiny and EmbeddingGemma 2
Memory	1933	OpenViking context server
Observability	6006	Phoenix
App	8668	The terminal style chat UI

Both models live in a single llama.cpp server that runs in router mode, so the request body decides which one answers. One CUDA context instead of two, which is what makes the card comfortable.

Run it

pixi install
docker/fetch-models.sh
docker/up.sh
docker/status.sh

Then open http://localhost:8668.

Labs

The lab instructions live in the repository, one page per experiment.

Lab index - how the labs are organised
Lab 1: Three truths - one question, three bodies of knowledge, watch them collide
Lab 2: Outside the training data - where the model stops being able to check itself

Repository

The labs - the experiments you can run
Project logo

The full source, the planning documents and the corpus pipelines live in the repository. This site is the readable part of it.