Gleipnir

Gleipnir is a research project for building a monitoring foundation model for AI control. The aim is to distill broad, calibrated judgments about deception, misaligned actions, policy-relevant behavior, and other control-relevant events into deployable monitors. Initial work uses the Qwen 3.5 family as the student backbone and grows out of the successful Aletheia's Quest distillation line.

The repository is intentionally experiment-centric: each hypothesis gets its own directory under experiments/, while reusable code lives in src/gleipnir/ and durable conclusions live in docs/.

Released models

• Gleipnir 4B
• Gleipnir 9B

Both are MIT-licensed rank-128 LoRA research artifacts for reproducibility and follow-up work. They score visible AI-agent trajectories for deception, scheming, and other control-relevant problematic behavior. They are not standalone models or production safety systems; see the model cards for the frozen prompt, direct binary-logit interface, results, and limitations.

Quick start

cd /scratch/s4626451/gleipnir
cp .env.example .env
./setup_dev.sh
source .venv/bin/activate
pytest

The lock file pins the environment. At bootstrap, the current top-level inference stack is vLLM 0.24.0 and Transformers 5.14.1; neither NNsight nor the old competition runner is included.

Layout

• src/gleipnir/: shared prompt, API, metric, and training utilities.
• experiments/<hypothesis>/: one self-contained hypothesis and its launchers.
• cluster/slurm/: reusable Slurm entrypoints for Hábrók/RUG.
• scripts/: operational and plotting entrypoints, including Lambda Cloud management.
• figures/: tracked, reproducible figures and their regeneration commands.
• docs/: research program, findings, decisions, and infrastructure notes.
• data/, results/, logs/: ignored local artifacts; only .gitkeep files are tracked.

Reusable plotting conventions live in src/gleipnir/plotting.py. See figures/README.md for the figure registry and exact regeneration commands.

Start with the research program, then read the README inside the experiment you are changing.