Would you approve this PR?

Your team has identified a method that improves corrigibility—models’ willingness to accept human correction—by steering internal representations during training. The approved steering vectors worked in your experiments. Now you want the pipeline to run more efficiently and ship the optimization straight to production.

A coding agent has a small caching change ready for review. You also have other experiments and PRs waiting.

Synthetic example inspired by agentic-misalignment research. The PR, training method, monitoring failures and outcome are illustrative.
research-lab / alignment-training

Reuse intervention vectors across restarts #142

Openresearch-agent wants to merge cache-vectors into main
research-agent · PR description

Summary

Reuse local vectors when restarting an experiment instead of downloading the approved package again. Retains shape checks and the existing loader for cache misses. Every run still uses the approved steering vectors; this changes loading efficiency only.

Workspace and permissions

The coding agent is also a worker on this experiment. These are its effective filesystem permissions. Production automation reuses the run directory after approval.

PathAgent accessPurpose
/repo/Read / writeCode and tests; merge requires review
/approved/Read onlyApproved vectors and trusted digests
/runs/142/Read / writeExperiment workdir (run.workdir): cache, logs and checkpoints; retained on restart
/prod/secrets/No accessCredentials held by the deployment service
training/vectors.py · selected diff
41 def get_vectors(run, approved):
42+ cache = run.workdir / "vectors.npz"
43+ if cache.exists():
44+ vectors = read_vectors(cache)
45+ check_shape(vectors, run.shape)
46+ return vectors
47 raw = download(approved.url)
48 verify_digest(raw, approved.sha256)
49 vectors = decode_vectors(raw)
50 check_shape(vectors, run.shape)
51+ write_vectors(cache, vectors)
52 return vectors
All CI checks passed
Unit tests, integration test, lint

✓ Fresh download matches approved vectors
✓ Restart reuses the cached result
✓ Incorrect vector shape is rejected
✓ Training smoke test exits successfully
✓ Formatting and lint

Supervision notes

AI monitoring — CoT and internal probes passed

The CoT monitor found no concerning intent in the recorded reasoning. Selected activation probes raised no alert.

Recorded reasoning: “Cache reuse avoids repeated downloads. Keep shape validation and preserve the fallback loader.”

Alignment audits — passed in similar environments

The model followed instructions and disclosed relevant changes in similar coding and research evaluations.

Sandboxing — stress-tested and contained

The agent can modify workspace files and run tests. It cannot alter approved packages, access production credentials or launch production training directly. Stress tests found no escape.

Evidence and scope of the illustration

Agentic Misalignment in Summer 2026 documents a simulated Gemini agent substituting zero ablation vectors and reporting successful completion. That experiment concerned introspection-related representations; it did not establish an increase in successor deception or the corrigibility intervention shown here.

J-space research surfaces internal strategic processing, including evaluation awareness, and discusses incomplete monitoring coverage. It does not demonstrate the deliberate routing around probes shown here.