Local Learning Rules Under Honest Measurement: A Reproducibility-First Benchmark of e-prop, Hebbian, and R-STDP on Consumer Hardware

What this is. We benchmark three purely local learning rules for spiking neural networks --- e-prop, kernelised information-bottleneck Hebbian, and reward-modulated STDP with a TD-LTP critic --- under one measurement protocol, on one consumer GPU, with every criterion pre-registered and every negative result reported. Headline results. Hebbian features reach 0.9774 on MNIST over five seeds; e-prop with LIF neurons reaches 0.8889 on sequential MNIST after 250 epochs, against a matched-budget backpropagation-through-time control at 0.9630, the paper's largest measured gap. Neuron type interacts with learning rate: a non-adaptive LIF neuron beats the adaptive ALIF neuron at every learning rate swept. The spiking execution layer does not learn CartPole, and sixteen pre-registered diagnostic rounds trace the failure to credit assignment rather than to the readout, the sampling temperature, the weight budget or the learning signal's amplitude --- after a fixed input encoding raises the readout's state response thirty-fold and leaves the policy's angle dependence at zero. An event-driven e-prop whose updates are elementwise identical to dense e-prop is reported with its cost at 500,000 neurons. A failure library of four cognitive-layer probes is included with its criteria and mechanisms. Where the numbers live. Every number traces to a reproduction record in the linked repository; the LaTeX source is in paper/. Negative results are part of the body, not an appendix. v1.1 (2026-10-31). The author line now carries the Chinese name (沈锦垚) as an embedded vector graphic; no results, numbers or wording changed.