Watch the mirage happen

This demo simulates context compression at different retention levels. As you lower the slider, more of the conversation is dropped. The "naive recency" policy keeps the newest messages regardless of importance. The "tropical guarded" policy uses mathematical contracts to protect the pivot message — the instruction that defines the task.

The validity mirage occurs when naive compression produces high raw validity scores (the AI sounds confident and coherent) while silently dropping the pivot. The AI is answering fluently — but answering the wrong question.

Naive Recency: Keeps the most recent messages, dropping older ones first. Simple and fast, but blind to task structure. When the pivot message is old enough, it gets dropped.

Tropical L2 Guarded: Uses tropical semiring algebra to identify and protect structurally important messages. The pivot and its k predecessors are never dropped, even under heavy compression. Trades slightly more tokens for guaranteed task preservation.