Inspired by llm-wiki, specialized for news world models.

What we keep from llm-wiki

Raw sources stay immutable. The app builds a compiled memory layer above them, and every important claim should remain traceable back to source articles.

What changes for news

The canonical objects are not wiki pages. They are event cards, actor notebooks, scenario forecasts, and review logs that can be recompiled as the story moves.

Why the model is intentionally simple

Instead of tracking every actor-to-actor edge, the model focuses on what each actor wants, what they fear, what scenario is becoming easier to imagine, and why.

1. Crawl and preserve sources

Articles are stored by topic, date, and outlet with source links, metadata, excerpts, and full text when available.

2. Compile event clusters

Related daily coverage is grouped into larger event phases, so repeated headlines become a readable story arc.

3. Build actor notebooks

For each important actor, the system keeps a topic-specific notebook: wanted future, unwanted future, hidden grievance, next move, and exaggerated inner voice.

4. Forecast scenarios

The current event arc and actor notebooks produce a small set of next scenes with likelihood, confidence, verifiability, and actor stances.

5. Recompile the world snapshot

The UI reads one compiled snapshot that joins source coverage, event phases, actor notebooks, scenarios, and review history.

Store the original reasoning

Each forecast keeps the reason it was made, not just the prediction label.

Review after the fact

When later coverage arrives, a review log records what happened, whether the scenario was right, and why the reasoning did or did not hold.

Feed reviews back into confidence

A scenario that was right for the right reason can gain confidence next time. Lucky hits, clear misses, and unknown misses lower confidence or stay unresolved.