Supplementary Code: Trauma as a Functional Kind

Anonymized supplementary repository for double-blind review. Authors: Anonymized for review.

This repository contains the four computational models referred to in the paper as the artificial realizer: neural-network demonstrations that the four functional profiles of maladaptive learning specified in the paper are realizable in a substrate whose mechanism is fully known and inspectable. The models are possibility proofs, not simulations of children or clinical instruments.

Contents

supplementary-repo/
├── trauma_models/               # Main package
│   ├── core/                    # Shared base classes, metrics, plotting, logging
│   ├── extreme_penalty/         # Model 1 — gradient cascade from a single extreme penalty
│   ├── noisy_signals/           # Model 2 — inconsistent labels -> weight/behavioral instability
│   ├── limited_dataset/         # Model 3 — few caregivers -> overfitting / generalization gap
│   └── catastrophic_forgetting/ # Model 4 — retraining destroys consolidated patterns
├── experiments/                 # YAML configs (exact hyperparameters per model)
├── tests/                       # Unit tests (architectures, reproducibility, statistics)
├── outputs/                     # Committed results backing the paper's numbers
│   ├── extreme_penalty_fixed/   # Model 1 run (data + figures)
│   ├── noisy_signals/           # Model 2 run
│   ├── limited_dataset/         # Model 3 — BOTH documented runs (see below)
│   ├── catastrophic_forgetting/ # Model 4 run
│   └── data/, figures/          # Additional Model 1 outputs
├── paper-figures/               # The exact PNGs embedded in the manuscript + Figure 4 script
├── MODEL_SPECIFICATIONS.md      # Mathematical formulations of all four models
├── requirements.txt             # Dependencies (pip)
├── requirements_frozen.txt      # Exact versions used for the committed runs
├── pyproject.toml               # Package metadata
└── run_all_experiments.sh       # Convenience runner for all four models

Requirements

• Python 3.8+ and PyTorch 2.0+
• All models run on standard CPU hardware — no GPU required (~2–4 GB RAM; the full experiment suite runs in roughly 5–20 minutes on a laptop CPU)

python -m venv venv
source venv/bin/activate      # Windows: venv\Scripts\activate
pip install -r requirements.txt

The Four Models

Model	Package	Failure mode	Developmental analogue
1	trauma_models.extreme_penalty	Single example with extreme loss penalty causes overcorrection in the boundary region around the penalized input	Single-event trauma generalizing to similar contexts
2	trauma_models.noisy_signals	Label noise produces weight variance, confidence collapse, behavioral inconsistency	Inconsistent caregiving / anxious-disorganized attachment
3	trauma_models.limited_dataset	Training on few synthetic "caregivers" yields a generalization gap on novel ones	Narrow early environment limiting generalization to new adults
4	trauma_models.catastrophic_forgetting	Aggressive retraining on a small corrective set destroys consolidated patterns; experience replay preserves them	Why safe therapeutic retraining is necessarily gradual

Mapping to Paper Figures, and Reproduction Commands

Note the off-by-one between graphic filenames and paper figure numbers: the file figure0_* is paper Figure 1, figure1_* is Figure 2, and so on. The exact PNGs used in the manuscript are committed in paper-figures/.

Paper figure	Graphic file	Generated by
Figure 1	figure0_four_categories_diagram.png	Conceptual diagram (not experiment output)
Figure 2	figure1_extreme_penalty_overcorrection.png	Model 1 experiment (below)
Figure 3	figure2_noisy_signals_instability.png	Model 2 experiment (below)
Figure 4	figure3_limited_dataset_overfitting.png	Model 3 statistical run + paper-figures/generate_figure4.py
Figure 5	figure4_catastrophic_forgetting_therapy.png	Model 4 experiment (below)

Run everything from the repository root.

Figure 2 — Model 1 (extreme penalty, seed 42, single run):

python -m trauma_models.extreme_penalty.experiment \
    --config experiments/extreme_penalty_sweep.yaml --output outputs/
# -> outputs/figures/extreme_penalty_overcorrection.png
# committed reference run: outputs/extreme_penalty_fixed/

Figure 3 — Model 2 (noisy signals, 10 runs per noise level, seeds 42–51):

python -m trauma_models.noisy_signals.experiment \
    --config experiments/noisy_signals_sweep.yaml --output outputs/
# -> outputs/noisy_signals/figures/noisy_signals_analysis.png

Figure 4 — Model 3 (limited dataset, statistical run: [2, 5, 10] caregivers × 10 trials, seeds 42–51):

python -m trauma_models.limited_dataset.statistical_significance
# -> outputs/limited_dataset/statistical_significance_analysis.{json,txt}
#    outputs/limited_dataset/figures/statistical_significance.png
python paper-figures/generate_figure4.py
# -> paper-figures/figure3_limited_dataset_overfitting.png (the manuscript figure)

Figure 5 — Model 4 (catastrophic forgetting, seed 42, single run):

python -m trauma_models.catastrophic_forgetting.experiment \
    --config experiments/catastrophic_forgetting_sweep.yaml --output outputs/
# -> outputs/catastrophic_forgetting/figures/catastrophic_forgetting_comparison.png

All four models at once:

./run_all_experiments.sh

Model 3 Has Two Documented Runs — Which One Backs Figure 4

outputs/limited_dataset/ contains results from two distinct experiments:

1. Statistical-significance run (statistical_significance_analysis.json / .txt): caregiver counts [2, 5, 10], 10 trials per condition (seeds 42–51, seed = base_seed + trial), 30 training runs total. This is the run reported in the paper and plotted in Figure 4. All Figure 4 numbers trace to it: generalization gap 0.0161 ± 0.0098 (2 caregivers) vs 0.0058 ± 0.0027 (10 caregivers), the pre-planned 2-vs-10 contrast t(18) = 3.20, p = 0.005, Cohen's d = 1.43, surviving Holm correction.

2. Single-seed sweep (limited_dataset_results.json): caregiver counts [2, 3, 4, 5, 6, 8, 10, 12, 15, 20], one run per condition (seed 42), produced by python -m trauma_models.limited_dataset.experiment. This finer-grained sweep is retained for completeness and exploratory inspection; because each condition has n = 1, it supports no error bars or tests and is not the source of any number in the paper.

Seed Conventions

All experiments use fixed seeds; identical seeds give identical numerical results (pixel-level PNG output may vary slightly across matplotlib versions).

Model	Seeding
1 — extreme_penalty	Single run, seed 42 (from experiments/extreme_penalty_sweep.yaml)
2 — noisy_signals	10 runs per noise level, seeds 42–51 (run_seed = 42 + run_idx)
3 — limited_dataset (statistical run)	10 trials per condition, seeds 42–51 (seed = 42 + trial)
3 — limited_dataset (single-seed sweep)	One run per condition, seed 42
4 — catastrophic_forgetting	Single run, seed 42

Tests

pytest tests/

The suite covers model architectures, gradient-cascade behavior, reproducibility (same seed → same results), and the Model 3 statistical pipeline. Tests were written against the pinned versions in requirements_frozen.txt (torch 2.9.0). One behavioral assertion (test_gradient_cascade.py::test_overcorrection_increases_with_penalty) depends on the exact PyTorch build's RNG stream and has been observed to fail under a different build while the remaining 25 tests pass; it exercises a short auxiliary training loop, not the committed results, which were produced with the frozen environment.

License

MIT (see LICENSE). Copyright holder anonymized for review.