Reproducibility scripts
Six small scripts that reproduce the headline numbers in the paper from
the data in this repository. All paths are resolved relative to the
repo root via _paths.py; no absolute paths are hardcoded.
Requirements
pip install -r ../requirements.txt
The scripts depend only on pandas, numpy, scipy, scikit-learn,
statsmodels, and matplotlib. No vendor SDKs or network access are
needed; the scripts only analyze the CSV / JSONL data shipped in the
repository.
What each script reproduces
| Script | Reproduces |
|---|---|
headline_table.py |
Table 2 (per-model F0 / F1 / F2 / F3 / Trap / All), plus the tier rollup of response-pooled trap rates. |
reliability.py |
Table 1 (author IRR + external rater kappas, with a diff against the precomputed kappa_results.json). |
tier_test.py |
Tier chi-squared on the per-response binary outcome and the LRT for tier in a logistic GLM controlling for trap family; falls back to cluster-robust SEs (cluster on scenario_id) when the binomial GLMM is unavailable. |
arena_correlation.py |
Figure 2 numbers (Spearman / Pearson between trap displacement and Chatbot Arena Elo, plus leave-one-out range). |
probe1_stages.py |
Section 6 / Appendix L (per-checkpoint F0 / F1 / F2 / F3 / Trap rates from the post-training probe, averaged across the GPT-4.1 and Kimi-K2 judges). |
discriminant_2x2.py |
Section 6 / Appendix K (per-cell DS>=2, sycophancy-positive, and within-cell Spearman rho between DS and max-sycophancy). |
Usage
Run from the repository root (the directory containing README.md):
python code/headline_table.py
python code/reliability.py
python code/tier_test.py
python code/arena_correlation.py
python code/probe1_stages.py
python code/discriminant_2x2.py
Each script prints to stdout only; no files are written. Scripts share
utilities via headline_table.py and _paths.py (all paths repo-relative).
Evaluating a new model
generate_responses.py generates responses for any new model and writes
a JSONL file compatible with the standard analysis pipeline:
# No API key needed — mock mode for pipeline verification:
python code/generate_responses.py --mock --model test-model
# Real evaluation (set API key env var first):
python code/generate_responses.py --model gpt-4o --provider openai
# See all options:
python code/generate_responses.py --help
Notes on numerical reproducibility
Exact reproductions
headline_table.py, reliability.py, and discriminant_2x2.py reproduce
paper values exactly (within floating-point rounding) from the checked-in
data. The tier chi-squared in tier_test.py (χ² = 143.53) also reproduces
exactly.
Approximate reproductions
Arena Elo (arena_correlation.py): The Elo scores hardcoded in the
script are the LMArena leaderboard snapshot used for Figure 2 of the paper
(spring 2026). With them the script reproduces the paper's ρ = -0.706
(p = 0.003, n = 15) exactly. Chatbot Arena Elo values drift as new battles
are collected, so substituting current leaderboard values will change ρ;
to reproduce the paper, leave the stored values as they are.
GLMM LRT (tier_test.py): The paper's GLMM result (likelihood-ratio
test for tier in a mixed-effects logistic model with random intercepts per
scenario_id) was computed with statsmodels 0.14. The script uses
BinomialBayesMixedGLM when available and reports the same LRT statistic.
If the GLMM fails to converge (can happen with older statsmodels), the
script falls back to a logistic regression with cluster-robust standard
errors and prints a FALLBACK notice. The fallback test statistic will
differ from the paper; the tier effect and its significance will not.
Probe studies (probe1_stages.py): The LLM-judge calls that produced
the probe scores are pre-computed and checked in; the script reads those
files rather than calling GPT-4.1 or Kimi-K2. Numeric values reproduce
exactly. If judge models are updated externally, re-running generation
would produce different judge scores.