EdDynamics result reproduction

This repository accompanies the preprint Reproducible macroscopic dynamics in a closed-loop human-AI learning system, available at https://doi.org/10.48550/arXiv.2608.30946.

Replace every /path/to/... placeholder before running a command.

Layout

s1_empirical_dynamics/
├── main/
├── scripts/
└── robustness/
s2_minimal_mechanism/
├── main/
├── scripts/
└── robustness/
s3_event_ssl/
├── main/
│   ├── controls/
│   └── structural_analyses/
├── scripts/
└── robustness/
s4_cross_analysis/
├── main/
├── scripts/
└── robustness/
s5_supplementary/
└── scripts/
reference_outputs/

See reference outputs for core numeric reports and supplementary report-generation scripts.

The s1–s5 directory prefixes are navigation labels. In contrast, stage* prefixes retained in code filenames and result directories denote the actual workflow order.

main/ contains primary experiments and downstream analyses, robustness/ contains additional robustness workflows and audits, and scripts/ contains result extraction and figure commands.

Dependency order

1. Generate the empirical outputs in s1_empirical_dynamics/main/.
2. Run both construction-matched null analyses and empirical sensitivity from s1_empirical_dynamics/robustness/.
3. Run the complete ordered mechanism sequence in s2_minimal_mechanism/main/.
4. Run primary Event-SSL, its controls, macro sufficiency and representation geometry using s3_event_ssl/README.md.
5. Run the five additional Event-SSL seeds, then the state-only and objective-control workflows.
6. Complete the remaining empirical and mechanism robustness workflows.
7. Run s4_cross_analysis/README.md.
8. Run the scripts/README.md inside each component when its stated inputs exist; run s5_supplementary/scripts/README.md last.

Resume a wrapper with the same result root when it provides completion markers. Do not combine outputs generated from different code copies in one result root.