Datasets:
rho float64 0 1 | phi_err float64 0 0.92 | W_exact float64 0.07 0.91 | union_bound float64 0.22 2.88 | bound_over_exact float64 3.04 3.43 | bound_informative bool 2
classes |
|---|---|---|---|---|---|
0 | 0 | 0.911 | 2.883 | 3.16 | false |
0.5 | 0.044 | 0.85 | 2.756 | 3.24 | false |
0.7 | 0.112 | 0.763 | 2.561 | 3.36 | false |
0.9 | 0.323 | 0.569 | 1.952 | 3.43 | false |
0.95 | 0.48 | 0.462 | 1.5 | 3.25 | false |
0.99 | 0.761 | 0.226 | 0.689 | 3.04 | true |
0.999 | 0.924 | 0.072 | 0.218 | 3.04 | true |
Verification Bandwidth Under Correlated Evaluators — derived results
Derived numerical results for the paper Verification Bandwidth Under Correlated Evaluators: What an Effective-Sample-Size Statistic Measures in an Acceptance Cascade.
- Paper concept DOI: 10.5281/zenodo.21891435
- Paper version DOI (v1.0.0): 10.5281/zenodo.21891436
- Code: github.com/spectralbranding/orgschema-papers/tree/main/verification-bandwidth/code
- This dataset's DOI: 10.57967/hf/9953
What this is, and what it is not
This is not an observational dataset, and it should not be cited as evidence. The paper collected no data. It is a theory-and-computation paper whose reported figures come from two places: arithmetic on summary statistics another study published, and seeded simulation of a stated model. This record is the output of that computation — the tables the paper prints, plus the captured stdout of every script that produced them.
Its purpose is auditability rather than reuse. A reader who wants to check a number in the paper against the code that produced it can diff this record instead of re-running anything; a reader who wants to re-run it can, from the repository above, with no network access and no key.
Contents
| File | What it holds |
|---|---|
tables/table1_bracket_width_vs_union_bound.csv |
Exact bracket width against its union bound across the correlation range, with the looseness factor and whether the bound is informative at all |
tables/table3_reported_neff_vs_formula.csv |
The published nine-judge panel's reported effective sample sizes beside the design-effect formula evaluated at its own reported error correlations |
tables/table4_bracket_surviving.csv |
Percentage of the zero-correlation bracket width surviving at that panel's error correlations, bracketed over three marginal error rates |
tables/table5_dimensional_ceiling_typical.csv |
The typical-case dimensional ceiling: exact limit and its square-root asymptotic, by state-space dimension |
tables/table6_transferable_share.csv |
Transferable share and accountable-signatory residual under a correlated receiving panel |
logs/*.log |
Captured stdout of all six scripts, including the seeded Monte Carlo tables (paper Tables 2 and A1) that are deliberately not re-derived as CSV |
figures/*.png |
The map from inspection geometry to error correlation, and geometric against error correlation |
Provenance and how to reproduce
Every value is produced by reproduce.sh in the repository above, which runs six scripts in dependency order. All fix SEED = 20260811 at file top and exit nonzero if any internal check fails. The whole pipeline runs in well under a minute and requires only Python 3.12 with numpy, scipy and matplotlib.
git clone https://github.com/spectralbranding/orgschema-papers
cd orgschema-papers/verification-bandwidth
./reproduce.sh
Tables 2 and A1 and the worst-case block of Table 5 are seeded Monte Carlo. They appear here only as captured stdout, not as CSV, because emitting them would require a second implementation of a seeded simulation — the exact drift the paper's reproducibility standard exists to prevent. The script is the ground truth for any value the paper calls computed.
Known limits of what these numbers mean
- Table 3 is arithmetic on another study's published figures, not a re-analysis of its panel. No raw judgements were obtained; the agreement it reports is between a formula and a published number.
- Table 4 is model-dependent and directional. The point inversion needs a marginal error rate the published record does not carry, so it is bracketed over three values rather than fixed, and no shared-difficulty correction is applied.
- Table 6 rows are illustrative combinations, not measurements of any organization, and are computed under one-dimensional inspection subspaces.
- The simulations assume isotropic deviations, rank-one inspection, and a miss-only error model with no false-alarm arm. Each is a stated scope condition of the paper, not a defect of this record.
Citation
Cite the paper, not this record:
Zharnikov, Dmitry (2026), Verification Bandwidth Under Correlated Evaluators: What an Effective-Sample-Size Statistic Measures in an Acceptance Cascade. Working Paper v1.0.0. DOI: 10.5281/zenodo.21891435
License
CC BY 4.0. The code that produced these results is MIT, at the repository above.
- Downloads last month
- -