Simulation data for "Internal-state criticality in Bayesian–inverse-Bayesian inference"

# Data archive — Internal-state criticality in Bayesian–inverse-Bayesian inference

**Paper.** *Internal-state criticality in Bayesian–inverse-Bayesian

inference*, K. Sasai and Y.-P. Gunji (Physical Review Research, submitted).

**Source repository.** <https://github.com/kazsasai/bayesian-inverse-bayesian-rps>

**DOI.** `10.5281/zenodo.20533918`.

---

## Contents

This deposit contains the raw simulation outputs underlying every data figure

of the paper. Three archives are provided: two cover the main simulation data

(*full reproducibility* vs *quick figure rebuild*), and one small archive holds

the reinforcement-learning baseline-control data:

| Archive | Size (compressed) | Contains | Use case |

|---|---|---|---|

| `paperA_data_full.tar.gz` | ~2.7 GB | Full simulation output tree (~16 GB uncompressed; 2713 files): all per-run JSONs and NPZs from `simulation/{reward_huge,nhand,reward_huge_v2,analyze_sharpness_plateau,reward}/data/` and `simulation_tie_mode_ablation/data/` | Independent re-analysis from raw outputs |

| `paperA_data_figure_only.tar.gz` | ~1.1 GB | The 163 specific JSON/NPZ files actually read by `build_all.py` (~1.7 GB uncompressed) | Rebuild figures only |

| `paperA_data_baseline_control.tar.gz` | ~21 MB | Pooled run-length arrays (`pnas_rl_comparison/data/baseline_dwells.npz`) for the RL-baseline control — WSLS, tabular Q-learning, and regret matching vs BIB; 40 seeds, T=2e5 — backing Fig. 4 (`fig_control_ab`) | Rebuild the RL-baseline control figure |

The two main archives preserve the relative-path layout so that extracting

either at `<repo>/data/` lets `build_all.py` find the data without further

configuration. The baseline-control archive instead carries the

`pnas_rl_comparison/data/…` path and extracts at the **repository root**.

See **Reproducing the figures** below.

Supporting files:

* `MANIFEST_canonical.txt` — the in-repo data manifest

(`BIB_Levy_v2/latex/figures/scripts/zenodo_data_manifest.txt`), listing each

data tree, the figure(s) it feeds, and the generating script.

* `figure_only_file_list.txt` — exhaustive 163-line list of relative paths

inside `paperA_data_figure_only.tar.gz`, captured by auditing every

`open()` call from a clean `build_all.py` run (and re-running with caches

cleared so that no precomputed intermediate hid raw-data references).

* `checksums.sha256` — SHA-256 of all three archives.

## Reproducing the figures

Both tarballs preserve the same layout, so the workflow is identical:

```bash

# 1. Clone the source repo

git clone https://github.com/kazsasai/bayesian-inverse-bayesian-rps.git

cd bayesian-inverse-bayesian-rps

# 2. Get the data: pick ONE archive

# (full = raw-output independent re-analysis;

# figure-only = just enough to rebuild figures)

mkdir -p data

tar xzf /path/to/paperA_data_figure_only.tar.gz -C data # OR _full

# 3. Install dependencies

pip install numpy matplotlib powerlaw

# 4. Rebuild figures

python BIB_Levy_v2/latex/figures/scripts/build_all.py

# (or run individual scripts: build_Fig3_universality.py, etc.)

```

Alternatively, point `PAPERA_DATA` at an extraction directory anywhere on disk:

```bash

tar xzf paperA_data_figure_only.tar.gz -C /scratch/papera_data

export PAPERA_DATA=/scratch/papera_data

python BIB_Levy_v2/latex/figures/scripts/build_all.py

```

`figdata.py` in the source repo searches `$PAPERA_DATA`, then `<repo>/data/`,

then the in-repo `simulation/` tree, in that order.

### RL-baseline control figure (Fig. 4)

`paperA_data_baseline_control.tar.gz` carries the `pnas_rl_comparison/data/…`

path, so extract it at the **repository root** (not `<repo>/data/`):

```bash

tar xzf /path/to/paperA_data_baseline_control.tar.gz -C bayesian-inverse-bayesian-rps

python pnas_si/figures/build_fig_control.py # -> fig_control_ab.{pdf,png}

```

The figure's BIB curves are read from the main data (the `reward_huge_*`

`durations_bib-*` JSONs in the full / figure-only archive, via `$PAPERA_DATA`);

the baseline curves come from the archive above. To regenerate the baseline

data from scratch instead (deterministic, ~minutes):

```bash

python pnas_rl_comparison/run_baseline_control.py # -> baseline_dwells.npz

python pnas_rl_comparison/analyze_baseline_control.py

```

## What `paperA_data_figure_only.tar.gz` excludes

* The 17 G of per-run / per-step JSONs in the data trees that no current

figure reads.

* Intermediate caches (`fig*_ccdf_cache.json`) — these are regenerated by

`build_Fig4_robustness.py` and `build_FigS2_nh_ccdf.py` on first run.

* The small bundled inputs already shipped with the GitHub repo at

`BIB_Levy_v2/latex/figures/scripts/data/` (`scheme_summary.csv`,

`bo_tournament_results.json`, `sigma_*_rs_bib-bib.json`,

`data_ivb_{equil,biased}.npz`). The build scripts read these straight from

the repo.

## Verifying integrity

```bash

shasum -a 256 -c checksums.sha256

```

## Citation

If you use these data, please cite both the paper (forthcoming) and this

Zenodo record. The repository's `README` is updated with the final citation

on publication.

## License

Data are released under CC-BY-4.0 (deposit metadata sets this on Zenodo).

Source code in the GitHub repository is under its own LICENSE file.