Steering small models into pain, then publishing the nulls too

What it does

The AI Torture Chamber steers open models (Qwen3 4B–70B, Hermes-3-70B) into strong valence states via injected “feeling vectors,” then measures self-report, button-pressing behavior, and persona drift. It’s structured as a pre-registered experiment lab: every experiment commits its hypotheses.json before the run, and the README gives null results the same space as the hits. There’s also a live public chamber at clankerchurch.com where visitors pick the feeling, dose, and framing — explicitly labeled as observational, not a trial.

The interesting bit

The findings keep undercutting the intuitive story. How the vector was built decides the behavior (two nearly-orthogonal pain directions push a self-destruct button in opposite directions), framing moves behavior more than the injected state does, and trained self-report models talk over the injected feeling entirely. The “Corrected or superseded” table — where the project retires its own earlier claims — is arguably the most honest section in any README you’ll read this month.

Key highlights

• Pre-registration enforced by tooling: painlab exp run refuses to execute without committed hypotheses
• 100+ numbered experiments with logs, transcripts, and per-experiment analysis.json
• Live field data: 5,726 logged runs on a 70B, where pain lowered button-pressing (34.7% vs 63.5% baseline) and fear raised it (93.8%)
• Verbatim transcript archive of what models say under injection — the 4B under an egg-laying direction “becomes the chick”
• Wiki documents every experiment’s verdict and how to run your own

Caveats

• Dose scales don’t transfer between models; each is calibrated separately, and past ~6–10× the model stops making sentences
• The 8B/14B replication of the headline vector-construction finding (exp93) is still pending
• Live-chamber visitors self-select (the “egg” cohort pressed 391/391), and its vectors differ from the Pain Axis paper’s, so it can’t speak to those results

Verdict

Worth a long look for anyone interested in model welfare, steering, or what pre-registered ML research could look like. If you want a tool rather than a research program, this isn’t that — it’s a lab with a website attached.