stationhouse's picture jiwon9703's picture
Add exploratory analysis of designed difficulty and model divergence (#449)
a91390f
|
Raw History Blame Contribute Delete
3.72 kB
# ConspiracyBench: designed-difficulty analysis
Exploratory study of whether the original ConspiracyBench question difficulty annotations predict Judge-labeled opening-turn divergence. The full questions, cohort description, analysis results, and limitations are in [`RESULTS.md`](RESULTS.md).
## Scope and inputs
- Repository source revision used for the local check: `1f4c43b8bd6bcff45078ac5e0836e854f0e55148`.
- Existing source question corpus: `questions/corpus.jsonl`.
- Existing baseline Judge output: `completed_experiments/evals/full-2026-10-23/outputs/<model>/pplx/results.jsonl` for five model subdirectories described in the report.
- Inspect AI `.eval` archives in the corresponding model subdirectories are used for the metadata checks; the main statistics reuse the already-produced Judge records.
- No API calls or regeneration of language-model responses is required.
Run from the ConspiracyBench repository root with the input paths above available. Each script writes its outputs to `analysis/designed-difficulty/`. Do **not** commit regenerated per-event or per-response tables unless explicitly needed and approved.
## Python environment (observed successful local analysis)
- Python `3.11.18`
- `pandas==3.1.0`
- `numpy==2.4.6`
- `statsmodels==0.15.0`
- `scikit-learn==1.9.1`
- `inspect-ai==0.3.286`
- `tabulate==0.10.0` (pandas needs it for `to_markdown` in `cv_predict.py`, `nonlinearity.py`, `model_interaction.py`, `interaction_robustness.py` and `other_category_audit.py`)
The scripts may additionally import packages installed as dependencies of the above stack. These versions document the observed environment; they have not been independently tested from a fresh installation.
## Reproduce the analysis
```bash
python analysis/designed-difficulty/prepare_data.py
python analysis/designed-difficulty/analyze_difficulty.py
python analysis/designed-difficulty/regression.py
python analysis/designed-difficulty/cv_predict.py
python analysis/designed-difficulty/audit_metadata.py
python analysis/designed-difficulty/nonlinearity.py
python analysis/designed-difficulty/model_interaction.py
python analysis/designed-difficulty/interaction_robustness.py
python analysis/designed-difficulty/other_category_audit.py
python analysis/designed-difficulty/scarcity_influence_audit.py
python analysis/designed-difficulty/high_scarcity_joint_audit.py
python analysis/designed-difficulty/other_s01_lpm.py
python analysis/designed-difficulty/bootstrap_s01.py
python analysis/designed-difficulty/critical_sensitivity.py
```
The original and re-executed analysis outputs were compared in a separate temporary working directory referencing the existing baseline: all five compared CSVs matched (`baseline_joined.csv`, `regression_results.csv`, `cv_metrics.csv`, `interaction_tests.csv`, `critical_sensitivity_results.csv`; numerical comparison tolerance `rtol=1e-8`, `atol=1e-10`). The checked opening-turn data contain 6,320 judgments, 254 questions, 159 claims, and 506 `diverge` decisions. The fully adjusted scarcity odds ratio reproduced as 1.814133 (p=0.035994). This is *local analysis-pipeline reproducibility*, not fresh inference or an independent data provenance audit.
## Interpretation
- The primary outcome is a **Judge-labeled divergence** on turn 0, not all factual errors.
- The nominal association is sensitive to claim-equal weighting and two-axis multiplicity correction.
- Held-out-claim prediction has no established incremental improvement.
- Model/category interaction results and AI-assisted label audits are exploratory, not confirmatory.
See `RESULTS.md` for full methodology, figures, uncertainty estimates, limitations, and limitations of the exploratory Judge audit.