spectralbranding/band-conditionality-tables

Band-Conditionality of a Reservoir Capacity Ranking

Derived result tables, run logs and figures for the methodological note Rank Order Under a Capacity Benchmark Is Conditional on the Observation Band and the Delay Horizon (Zharnikov 2026, concept DOI 10.5281/zenodo.22206844).

The configs: block in this card's frontmatter is load-bearing, not decoration. Each table below has a different schema, so Hugging Face's default heuristics try to concatenate them into one split and the viewer fails with a cast error. One config per CSV is what makes the viewer work. Adding a table means adding its config.

Every file here is generated output, not hand-authored. The code that produces it is open at https://github.com/spectralbranding/band-ordering-experiment; running ./reproduce.sh from a clean clone regenerates every file in this dataset byte-identically, with no network access and no provider key.

What the study measured

Three leaky echo-state networks differing only in leak rate, and three damped Duffing oscillator banks differing only in natural frequency, were run across a band grid under two manipulations:

• Arm A (readout band) — native drive, every k-th state retained. The only manipulation available to a re-analyst of an archived recording.
• Arm B (drive band) — drive and readout coarsened together by zero-order hold. A genuine band change.

Sample count is held fixed at 350 across bands so that band is not confounded with estimator bias. Ten seeds per variant; intervals are t-based at 95%, df = 9. Seed 20260830 throughout.

Files

File	Contents
band_ordering_rows.csv	Main grid: capacity per variant, per band, per seed, both arms
saturation_rows.csv	Saturation control at fixed native band — the check that excluded the physical system from every ordering
e1_architecture_rows.csv	The same grid on the oscillator-bank family
e2_max_delay_rows.csv	Delay horizon swept 2–32 at two sample budgets, with basis size and samples-per-basis-term per cell
e3_block_resampling_rows.csv	Moving-block resampling of the single physical recording, and the same procedure applied to the simulated variants as validation
e4_aliasing_rows.csv	Twelve log-spaced frequencies × eight bands × both arms — the sweep that refuted the aliasing account
e6_charc_axes_rows.csv	Kernel rank, generalisation rank, memory capacity and capacity across the band grid, leaking-rate family
e6b_charc_axes_oscillator_rows.csv	The same four axes on the oscillator-bank family
e7_single_stream_rank_rows.csv	The second rank protocol: kernel and generalisation rank computed from one stream sampled at every observed timepoint, across the band grid, both arms, under both a fixed observed-state count and the fixed-drive confound
e7b_single_stream_rank_oscillator_rows.csv	The same second protocol on the oscillator-bank family, every setting but the substrate held identical to the row above; carries the tau_drive_steps column the leaking-rate table does not
logs/	Run logs for the two behaviour-space sweeps
figures/	The three figures as published in the note

What is NOT here, and why

The physical recording the study re-analyses is a public test file distributed with the benchmarking library used and is not redistributed in this dataset. reproduce.sh fetches it from its own source. Nothing else is withheld.

Reading the tables honestly

Five cautions, all stated in the note itself:

1. The physical system is excluded from every ordering claim. At the affordable sample budget it has not saturated the estimator, so its position would report starvation rather than capacity. Its rows are present so the exclusion can be checked, not so it can be ranked.
2. The aliasing account is retracted. e4_aliasing_rows.csv is the sweep that refuted it. What replaces it is a description — a graded, thresholdless timescale mismatch — not a mechanism.
3. Both e7_* tables contain a declared confound arm. Rows with hold = drive hold the drive length fixed, so the observed-state count falls as 1/k and rank tracks the column count rather than the band. They are published so the confound can be seen, not so it can be read as a result. Use hold = states.
4. The kernel-rank ceiling in e7_single_stream_rank_rows.csv does not generalise. On the leaking-rate family kernel rank sits at the node count in every cell, so it discriminates nothing there. On the oscillator family it does not: the slowest variant sits below the node count at the narrowest band and rises to it under a readout-only band change. That was predicted otherwise before the run, and the failed prediction is recorded in the note rather than absorbed.
5. The objects compared are parameterised variants within one architecture family, not different substrates. The substrate column name is the original data schema and is retained so the tables round-trip against the published code; it is not a claim about materials.

Citation

Zharnikov, D. (2026). Rank Order Under a Capacity Benchmark Is Conditional on the Observation Band and the Delay Horizon. Zenodo. https://doi.org/10.5281/zenodo.22206844

License

CC BY 4.0 for the data. The generating code is MIT, at the repository linked above.