EC-SPP: An Equivalence-Class Benchmark for Signal-Probability Prediction

Learned circuit encoders are graph neural networks that read a logic circuit and predict properties of its gates. The existing models are compared almost entirely through one number, the mean absolute error (MAE) of signal-probability prediction on a released And-Inverter-Graph dataset. We examine the released artifacts and find two properties that limit what comparisons at this scale can resolve. First, the labels are Monte-Carlo estimates carrying statistical uncertainties, and a model that outputs the exact probabilities still scores $0.00309$, one third of the best reported MAE. This creates an unavoidable error we call the label floor. Second, the benchmark scores one frozen netlist per circuit, while synthesis presents the same function in many equivalent encodings. Re-encoding a circuit, without changing anything it computes, moves the released DeepGate2 model's predictions by three times the label floor, and by eighteen times under a stress-test rewrite that structural hashing cannot remove. We package these measurements as EC-SPP, a benchmark of certified equivalence classes with labels at a fifteen-fold finer floor and consistency across encodings as a reported metric. On its baseline, evaluated deterministically, the choice of representative moves DeepGate2's score by fourteen times the top-two gap at the median circuit, and standardizing the whole benchmark on a single synthesis recipe shifts the aggregate score by six times that gap.