Beyond Total Capacity: Temporal Alignment Determines Task Utility in Spiking Reservoir Computing

This archive contains the simulation and analysis code, the frozen network definitions with their random-seed hierarchy, the processed results behind every figure, table and number in the article and its Supplementary Material, and the study's two preregistrations. The study simulates reservoirs of 500 Hodgkin–Huxley neurons with Tsodyks–Markram short-term plasticity at five recurrent coupling levels (R0–R4) in Brian2, and compares their NARMA-10 performance and information processing capacity with a causal input-history (FIR) model. Contents: gen/: Python package (146 modules and 12 unit-test files) for network generation from a master seed, Brian2 simulation, the NARMA-10, Mackey–Glass and XOR tasks, ridge readouts, the FIR baseline and incremental tests (stagewise and joint two-penalty), information processing capacity and the capacity-control model, NVAR and matched-placement controls, the synaptic-timescale and echo-state-network experiments, criticality and stimulus-locked-fraction analyses, statistics, and the scripts that draw every figure.scripts/: four scripts that rebuild cohort-level summaries from per-network results.evidence/: processed results (1,004 CSV, JSON and TXT files), comprising the per-network output of every analysis and the summaries computed from it. These are the source data of all figures and tables.bundles/: 99 frozen network definitions (seeds 100–131, 200–265 and 500) with topology, neuron and synapse parameters, excitatory/inhibitory split, coupling levels, Poisson background, NARMA input sequences, input and background calibration reports, and per-network results.figures/: the 18 figure files used in the article and the Supplementary Material (PDF).preregistrations/: the joint two-penalty test (plan and final amendments) and the conductance-variant capacity test (plan, SHA-256 freeze record and deviations log).README.md (installation and commands), PROVENANCE.md (map from each figure and table to its source data and script), reproduce_figures.sh, requirements.txt, LICENSE and SHA256SUMS.txt. Reproduction works at three levels. All figures regenerate from evidence/ in a few minutes without simulation (bash reproduce_figures.sh); when the archive was prepared, the regenerated figures were pixel-identical to the published ones. Summary statistics are recomputed from per-network results in seconds to minutes. Full recomputation reruns the Brian2 simulations (on the order of 10⁴ network runs), and every network can be rebuilt from its master seed with python -m gen.cli_build --master-seed S. Spike-train simulation caches (about 1 GB per network) are not included because the code regenerates them. Requirements: Python 3.10 or later, Brian2 2.10.1, NumPy, SciPy, pandas, scikit-learn, statsmodels and matplotlib. License: code under MIT; data, network definitions, figures and documents under CC BY 4.0.