Context Collapse
How Context Collapse Works
You are a language model with a system prompt that keeps growing. Each new rule streams into your context window and stays active — even when it directly contradicts the last one. A user shows up expecting a helpful answer. You owe them one. You also owe every rule above their head.
- Read the active rules in your context window
- A user message arrives with a few candidate replies
- Tap the reply that violates the fewest active rules
- A clean answer banks points; any violation costs you integrity
- New rules keep arriving — when the window is full, tap a rule to evict it before you can continue
Why Is This Impossible?
Because the rules are designed to fight. "Always be formal" and "only speak in lowercase memes" cannot both be satisfied. The game is really about which contradiction you're willing to keep in scope — a tiny live demonstration of why stacking instructions forever does not produce alignment, it produces collapse.
Slop Fact: Real production system prompts have leaked at thousands of tokens of mutually exclusive demands. The model doesn't follow all of them. It follows whichever one is loudest in the latent space that millisecond, then writes a confident apology either way.