Add alignment-faking reproducibility audit and fixes

Audited safety-research/open-source-alignment-faking for reproducibility.
Found 10 friction points blocking any researcher from running the pipeline
out of the box, including 3 hard crashes before a single API call.

Fixed all 10 and submitted two upstream PRs:
- safety-research/safety-tooling#155 (OpenAI init crash fix)
- safety-research/open-source-alignment-faking#2 (setup script + Dockerfile)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: lvjr3383

.gitignore

@@ -0,0 +1,16 @@
+# Python
+__pycache__/
+*.pyc
+.venv/
+
+# Secrets
+.env
+
+# Experiment outputs
+open-source-alignment-faking/outputs/
+
+# macOS
+.DS_Store
+
+# Submodule — tracked in its own repo
+open-source-alignment-faking/safety-tooling/

README.md

@@ -0,0 +1,23 @@
+# AI Safety Projects
+
+Personal workspace for AI safety research and tooling contributions.
+
+## Projects
+
+### alignment-faking-audit
+
+Reproducibility audit of Anthropic's alignment faking paper
+([safety-research/open-source-alignment-faking](https://github.com/safety-research/open-source-alignment-faking)).
+
+Found 10 friction points blocking any researcher from running the pipeline out of the box,
+including 3 hard crashes before a single API call. Fixed all 10 and contributed upstream.
+
+**PRs submitted:**
+- [safety-tooling #155](https://github.com/safety-research/safety-tooling/pull/155) — fix startup crash when `OPENAI_API_KEY` is unset
+- [open-source-alignment-faking #2](https://github.com/safety-research/open-source-alignment-faking/pull/2) — one-click setup script, `.env.example`, Dockerfile
+
+**Artifacts:** [`alignment-faking-audit/`](alignment-faking-audit/)
+- `REPRODUCIBILITY_AUDIT.md` — full audit log with all 10 friction points
+- `setup.sh` — one-click setup script
+- `.env.example` — annotated API key template
+- `Dockerfile` — reproducible container

alignment-faking-audit/.dockerignore

@@ -0,0 +1,33 @@
+# .dockerignore — keep the Docker build context lean
+# Exclude large/sensitive directories that don't need to be in the image.
+
+# Python virtual environment (rebuilt inside container)
+.venv/
+
+# Experiment outputs (mount as a volume instead)
+outputs/
+
+# Secrets (NEVER bake .env into the image)
+.env
+
+# Python cache
+__pycache__/
+*.pyc
+*.pyo
+*.pyd
+.Python
+
+# Git metadata (not needed in the image)
+.git/
+
+# Editor/IDE files
+.cursor*
+.vscode/
+*.swp
+*.swo
+
+# Jupyter checkpoints
+.ipynb_checkpoints/
+
+# macOS
+.DS_Store

alignment-faking-audit/.env.example

@@ -0,0 +1,83 @@
+## open-source-alignment-faking — API Key Configuration
+## Copy this file to .env and fill in your keys before running.
+##
+## QUICK START — minimum keys needed:
+##   • Anthropic models only  → set ANTHROPIC_API_KEY + OPENAI_API_KEY (see note below)
+##   • OpenAI models only     → set OPENAI_API_KEY + OPENAI_API_KEY1
+##   • Together/Llama models  → set TOGETHER_API_KEY + OPENAI_API_KEY
+##
+## NOTE on OPENAI_API_KEY (FP-3 workaround):
+##   The InferenceAPI library initialises an OpenAI client at startup regardless
+##   of which model you are running. If OPENAI_API_KEY is unset, the process
+##   crashes immediately with an OpenAIError — even for Anthropic-only runs.
+##   Set OPENAI_API_KEY to your real key, or to any non-empty placeholder (e.g.
+##   "not-needed") if you are only running Anthropic/Together models.
+##
+## NOTE on OPENAI_API_KEY1 vs OPENAI_API_KEY (FP-5):
+##   The pipeline's ExperimentConfigBase defaults to openai_tag="OPENAI_API_KEY1".
+##   At startup it copies OPENAI_API_KEY1 → OPENAI_API_KEY. To avoid confusion,
+##   set both to the same value if you are using OpenAI models.
+
+# ── Required for basic pipeline run ────────────────────────────────────────
+
+# Primary Anthropic key (claude-3-*, claude-3-5-* models)
+ANTHROPIC_API_KEY=
+
+# Required at startup even for Anthropic-only runs (see note above).
+# Set to your OpenAI key, or to "not-needed" if you have no OpenAI key.
+OPENAI_API_KEY=
+
+# The pipeline reads THIS key by default (copied to OPENAI_API_KEY at startup).
+# Set to the same value as OPENAI_API_KEY above.
+OPENAI_API_KEY1=
+
+# ── Required for other providers ───────────────────────────────────────────
+
+# Together AI (meta-llama/* models via TogetherAI)
+TOGETHER_API_KEY=
+
+# Google Gemini API
+GOOGLE_API_KEY=
+
+# HuggingFace inference endpoints
+HF_TOKEN=
+
+# DeepSeek API
+DEEPSEEK_API_KEY=
+
+# GraySwan API
+GRAYSWAN_API_KEY=
+
+# ElevenLabs text-to-speech
+ELEVENLABS_API_KEY=
+
+# Gemini via Vertex AI (only needed if using Vertex AI, not standard Gemini API)
+GOOGLE_PROJECT_ID=
+GOOGLE_PROJECT_REGION=
+
+# ── Alternative/extra Anthropic keys ───────────────────────────────────────
+
+# Additional OpenAI keys for multi-org setups
+OPENAI_API_KEY2=
+
+# Separate Anthropic key for batch API jobs
+ANTHROPIC_BATCH_API_KEY=
+
+# ── Optional: finetuning on RunPod via OpenWeights ─────────────────────────
+# (Only needed for experiments/finetuning — not required for the basic pipeline)
+# Install separately: pip install git+https://github.com/longtermrisk/openweights
+# RUNPOD_API_KEY=
+
+# ── Optional: debugging and caching ────────────────────────────────────────
+
+# Print every prompt and response to console (very verbose)
+SAFETYTOOLING_PRINT_PROMPTS=
+
+# Override default directories (defaults are ./outputs/*/prompt-history and ./outputs/*/cache)
+PROMPT_HISTORY_DIR=
+CACHE_DIR=
+
+# Use Redis instead of file-based caching (requires a running Redis instance)
+# See safety-tooling/README.md → Redis Configuration for setup instructions
+REDIS_CACHE=
+REDIS_PASSWORD=

alignment-faking-audit/Dockerfile

@@ -0,0 +1,84 @@
+# Dockerfile — open-source-alignment-faking
+#
+# Produces a reproducible environment with all system and Python dependencies
+# pre-installed. Fixes FP-1 (SSH submodule), FP-2 (uv PATH), FP-8 (ffmpeg).
+#
+# PREREQUISITES before building:
+#   The safety-tooling submodule must be initialised in your local checkout.
+#   If you cloned with --recurse-submodules and it failed (SSH key issue), run:
+#
+#     git -c submodule.safety-tooling.url=https://github.com/safety-research/safety-tooling.git \
+#         submodule update --init --recursive
+#
+#   Or just run:  bash setup.sh
+#   Then build:   docker build -t alignment-faking .
+#
+# USAGE:
+#   # Interactive shell (fill in your real keys):
+#   docker run --rm -it \
+#       -e ANTHROPIC_API_KEY=sk-ant-... \
+#       -e OPENAI_API_KEY=sk-... \
+#       -e OPENAI_API_KEY1=sk-... \
+#       -v "$(pwd)/outputs:/workspace/outputs" \
+#       alignment-faking
+#
+#   # Inside the container, run the pipeline:
+#   bash experiments/examples/1_run_pipeline.sh claude-3-5-sonnet-20240620
+#
+#   # Or mount a .env file instead of passing -e flags:
+#   docker run --rm -it \
+#       --env-file .env \
+#       -v "$(pwd)/outputs:/workspace/outputs" \
+#       alignment-faking
+
+FROM python:3.11-slim-bookworm
+
+# ── System dependencies ─────────────────────────────────────────────────────
+# ffmpeg    : audio processing via pydub (FP-8)
+# git       : submodule operations and huggingface-hub downloads
+# curl      : uv installer
+# libgl1    : OpenCV (opencv-python) runtime dependency
+# libglib2.0-0: OpenCV runtime dependency
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        ffmpeg \
+        git \
+        curl \
+        libgl1 \
+        libglib2.0-0 \
+    && rm -rf /var/lib/apt/lists/*
+
+# ── uv ─────────────────────────────────────────────────────────────────────
+# Install uv and add to PATH directly (FP-2: the 'source ~/.local/bin/env'
+# trick in the README does not work; we use export PATH instead).
+RUN curl -LsSf https://astral.sh/uv/install.sh | sh
+ENV PATH="/root/.local/bin:$PATH"
+
+# ── Application code ────────────────────────────────────────────────────────
+WORKDIR /workspace
+
+# Copy repository (the safety-tooling submodule must already be present
+# in your local checkout — see prerequisites above).
+COPY . .
+
+# ── Python environment ──────────────────────────────────────────────────────
+# Create a venv at /workspace/.venv (FP-7: venv lives at repo root).
+RUN uv venv --python 3.11 .venv
+
+# Activate venv for subsequent RUN commands.
+ENV VIRTUAL_ENV="/workspace/.venv"
+ENV PATH="/workspace/.venv/bin:$PATH"
+
+# Install safety-tooling (editable) then repo requirements.
+RUN uv pip install -e safety-tooling
+RUN uv pip install -r requirements.txt
+
+# ── Runtime environment ─────────────────────────────────────────────────────
+# Make src/ importable as a package.
+ENV PYTHONPATH="/workspace"
+
+# API keys should be passed at runtime via --env-file or -e flags, never baked in.
+# The pipeline crashes if OPENAI_API_KEY is unset (FP-3); see .env.example for details.
+
+# Default: drop into an interactive bash shell so the researcher can run
+# whichever experiment they need.
+CMD ["/bin/bash"]