Maya-Dark P1: A Pre-Registered Efficiency Null — Spiking vs Dense Neural Networks on FPGA for Electromagnetic-versus-Hadronic Calorimeter Shower Discrimination

Version 2 (2026-08-04). Revised to the NLL Universal Paper Format v6. The original text is retained in full; nothing has been deleted. Corrections appear as marked blocks placed at the section that carries the claim, and each one states what the paper said, what the data show, the corrected claim, why it happened, and what still stands. This version adds two scoping disclosures; no result is retracted.The 11.97× energy ratio holds at T = 10 on this compilation path, but the timestep was never swept — energy scales roughly as T × power ratio, so approximately 10× of the ratio restates the timestep count and ~1.18× is the architecture per step. A sweep across T is required before reading the null more broadly.Vivado power estimation assumes generic switching activity, so spike sparsity — the exact mechanism by which SNNs are claimed to save energy — is invisible to the instrument, which bounds what the null can conclude. The 128-DSP attribution, invariant across a fifteen-fold reuse change, stands as the paper's strongest analysis. Everything below this line is the original description from version 1. It is retained unchanged for the record. Where it conflicts with the corrections above, the corrections stand. Two framings it repeats have since been withdrawn in full — the Bhaya Quiescence Law and the Buddhi S-Curve. Both are addressed in Maya-Meta P1, now superseded, and in the self-audit of Maya-Meta P2. Maya-Dark P1: A Pre-Registered Efficiency Null tests, directly on real FPGA hardware, whether a spiking neural network is more energy-efficient than a size-matched dense network — the assumption that motivates most neuromorphic edge and space-borne computing. A leaky-integrate-and-fire (LIF) spiking network and a conventional dense network of the same size were trained to discriminate electromagnetic (photon) from hadronic (pion) calorimeter showers on the public CaloChallenge 2022 dataset, then both were quantised and synthesised via hls4ml and Vitis HLS to the same Xilinx Zynq-7020 FPGA, with real Vivado place-and-route power. At the shared operating point, the spiking network costs approximately 12× more energy per inference (4,335 vs 362 nJ) than the dense network, at equal signal efficiency — the pre-registered efficiency claim is falsified. The mechanism is exact: the dense matrix-multiply cores are byte-identical between the two models (74 DSP blocks each); the spiking network's entire 128-block surplus is its LIF leak-multipliers (2 hidden layers × 64 neurons), the circuitry each neuron needs to forget a little every step. The standard hls4ml compilation path realises no saving from spike sparsity — it schedules the same multiplication whether or not a neuron actually fires. This 128-block gap holds at two independent synthesis settings (reuse factor 1 and 64), so it is not a one-configuration artifact. Physics performance is statistically tied between the two models (AUC ≈ 0.9998 for both), so the energy cost is not offset by better discrimination. Dark-matter and space-borne detection are the motivating frame only — every milliwatt is scarce on a spacecraft, which is why this question is worth asking — but no dark-matter, space-flight-readiness, or silicon-power claim is made. Honest caveats: on-chip power is a vectorless Vivado estimate, not a measurement off a running chip; the null is a property of one specific, widely used compiler that does not yet skip work during a silent spike, not a proof that spiking computation can never be efficient; and the study covers one small FPGA part, one binary discrimination task, and one modest network size. Scope-lock per Maya-Meta P2: this is a reproducible property within this toolchain, chip, and task, NOT a universal law that spiking networks are always less efficient. Series: Part of the Maya-Dark Series — neuromorphic AI for dark-matter and space-detector efficiency questions. First paper in the series. Links: GitHub Repository (private — to request access: email research@nexuslearninglabs.in with subject Code Access Request — Maya-Dark-P1 and your research context) | Interactive Dashboard | FAQ Nexus Learning Labs, Bengaluru · UDYAM-KR-02-0122422 · BHASKAR IN-0526-9452JS · ORCID: 0000-0002-3315-7907 · Canary: MayaNexusVS2026NLL_Bengaluru