My calculator is a transformer

radarsat1 shows an RPN interpreter compiled into Transformer weights; “2 3 + 2 *” returns 10. The residual stream acts as registers, attention weights are compiler-calculated, while nonlinear MLP logic is still trained. The prototype is 1.1 GB; the key point is calculable attention weights, not a practical calculator.

Why it matters: HKR-H comes from the counterintuitive title; HKR-K has a reproducible input, weight-construction mechanism, and 1.1GB figure. HKR-R is real but niche, so this stays just above featured threshold, below 78.