kuilt-warp-ml/us.tractat.kuilt.warp/ReferenceTrainer/step

step

fun step(weights: List<Double>, examples: List<Pair<Double, Double>>, learnRate: Double): List<Double>

One GD step. examples are (x, y) pairs. Returns the updated [w0', b'].

Operation order is load-bearing: it is replicated verbatim in fedavg_train.wat, so the two produce bit-identical f64 results on the JVM.

Throws

IllegalArgumentException

if examples is empty. The mean-squared-error scale is 2 / |examples|, so an empty batch would divide by zero and return [NaN, NaN] — a NaN weight then dominates FedAvg's slot ordering and poisons the merged model, so fail loud rather than emit a plausible-looking wrong answer.