Algorithmic Intent Leakage in Transpiled Quantum Circuits: An Initial Reverse-Engineering Study

We study whether coarse algorithmic intent can be inferred from final backend-transpiled quantum circuits. The receiver observes only the optimized physical circuit and has no access to the source program, logical-qubit layout, routing history, or intermediate compiler representations. We construct a controlled dataset spanning seven intent classes and transpile circuits to Qiskit's FakeNighthawk backend for 24 and 48 qubits at optimization levels 0--3. Each circuit is represented by ten normalized features capturing gate composition, temporal two-qubit activity, and spatial hardware-load structure. A Random Forest classifier achieves high accuracy and macro-F1 when trained and tested within the same optimization-level or circuit-size regime. Cross-setting transfer is weaker, except between optimization levels 2 and 3, which show strong mutual generalization. These results indicate that final transpiled circuits retain measurable intent-related structure, but that this leakage depends on transpiler optimization level and circuit-size regime.