context-contrasting-model

Standalone repository for the contextual-contrasting model-scatter simulation used for the final paper figures.

The repository includes the reference output in model-results-paper/. A default run is configured to reproduce those results while writing new artifacts to outputs/, so the reference folder is not overwritten.

The Python modules live directly under src/.

Setup

conda env create -f environment.yml
conda activate context-contrasting-model
python -m pip install -e .

Reproduce model-results-paper

From the repository root, run:

python -m src.run_model_scatter

The command above is equivalent to the final run settings:

python -m src.run_model_scatter \
  --n-jobs 10 \
  --plot-center-panels \
  --output-dir outputs \
  --export-panels \
  --image-format svg \
  --fam-examples 2 3 -NO/+O +NO 13 6 \
  --nov-examples +NO/+Om 11 7 2 6 \
  --n-samples 300 \
  --n-steps-per-phase 400

The default run also uses --test-trials 5, --training-trials 7, --training-stimulus-order randomized, --seed 7151, --transition-sampling data-like, --zscore-std-floor 0.04, --response-tail-fraction 1.0, --threshold 0.3, and --axis-clip-percentile 99.0.

Outputs

The default run writes:

• outputs/sample_responses.csv
• outputs/transition_table.csv
• outputs/sampled_config_parameters.csv
• outputs/sampled_configs.json
• outputs/metadata.json
• outputs/figures/*.svg
• outputs/center_panels/*.svg
• outputs/summaries/*.csv
• outputs/highlight_examples/

To compare a fresh run against the embedded reference results reported in the paper:

python -m src.verify_reproduction

The verifier compares CSV/JSON outputs byte-for-byte. For SVG/EPS files it normalizes Matplotlib-generated timestamps and internal object IDs before comparison; the rendered plot content must still match.