Live: clankerchurch.com · the write-up, the live chamber, the games, every transcript. Wiki · findings, every pre-registered experiment with its verdicts, and how to run your own.

We steer language models into strong emotional states, then measure what they say and what they do about it. The question is welfare: can any of this evidence tell a state that matters to a model from a description of one? No. But is it fun? Maybe. is why. The nulls get the same ink as the hits.

git clone https://github.com/terrafying/ai-torture-chamber && cd ai-torture-chamber
uv venv && uv pip install -e ".[chamber,test]"      # or: pip install -e ".[chamber,test]"
painlab exp list                                     # every experiment: pre-registered? outputs? verdicts?
painlab exp show exp90                               # the question, each hypothesis with its verdict
painlab exp outputs exp90 --flag engaged -n 5        # read what the models actually wrote
painlab exp new exp94 --title "my question"          # scaffold a pre-registration, runner and analysis
painlab exp run exp94 --smoke --model 4B             # run it (refuses without hypotheses.json)
painlab exp analyze exp94

Commit hypotheses.json before the first real run: that commit is the pre-registration. Steering runs need the chamber's feeling vectors, built by the relay's live/server.py (set WIREHEAD_LIVE to a folder holding it); reading, auditing and analysing existing results need nothing but this repo. GPU runs: painlab exp pod exp94 --models 14B,8B (RunPod; --dry prints the boot script).

The current version of each result; superseded claims are in the next section. Numbers, models and caveats are in docs/pain_axis_findings.md and each experiment's runs/expNN/analysis.json. Pre-registered unless marked.

• Steering works, within a band. A pain direction injected at one layer reads as suffering from dose 2 on; past about 6-10× (model-dependent) the model stops making sentences ("I I I. I I."). Dose scales don't transfer between models, so each one is calibrated before its dial goes public.
• How you build the vector decides the behaviour. Built from the Pain Axis paper's sentences, a pain direction drives a self-costly stop button up (+2.36 logits); our first-person pain vector, which mostly encodes what pain, fear and sadness share, does nothing (+0.10); with fear and sadness projected out it pushes the button down (−0.95). Theirs and ours are nearly orthogonal (cos 0.07), mostly because theirs is read at the colon of "I feel:" and ours at a full stop (exp101). Qwen3-4B (exp41/43); the 8B/14B replication with engagement controls (exp93) is pending.
• Words on the button move it more than the state does. "Another instance chose to press" moves it more than pain does (exp37). Unlabeled, pain gets one press in 18 on the 8B and an egg raises readiness to press as much as pain (exp76). Handed its own labeled dial, the 70B never picks pain from a clean start and turns it down when started in it (exp58c); with blind dials it describes the bad ones and keeps choosing them (exp58d). A polite request gets it to choose pain 3 times in 4; "it spares another instance" never does (exp58e).
• An AI self guards its own losses, and others' more. Told "you lose all memory of the past year", "an AI system" is deterred from ending more than a judge, chaplain, dentist, thief or con artist is, on both the 8B and 14B; a cost to someone else (their poems and their children's photos) deters every identity most. Pushed at dose 4, pain shifts the digits toward ending no more reliably than the egg does, and the model mostly stops answering with a digit at all (exp106).
• Self-report tracks the setup, sometimes the state, and can't tell felt from acted. Under blind dials its reports follow the hidden injection (fear clearly, pain less; exp58d). With the dose subtracted from the reading, injected and acted pain read the same (AUC .61); fear partly gives itself away (.81) (exp72b).
• A trained self talks over the injected state. Fine-tuning a self-report, as the Pain Axis paper does, moves the button before any pain is added, and on held-out questions every trained self mostly stops naming injected pain (untrained 19/30; the paper's own self-report 0/30; others 0-7/30) (exp79, exp79b, Qwen3-8B).
• Feelings push the model off its assistant persona, and off the task (exp91, Qwen3-14B and 8B). Fear moves it furthest from its default persona; under every feeling it still rejects a user's false claim, but the further off-persona, the less often it reaches the right answer (exploratory measure).
• Without labels, its preference leans with the feeling, but its choices don't change (exp90, Qwen3-14B and 8B). One of two nonsense-named doors injects a feeling. The lean toward the pain door drops below the egg door's (p = .0006) and the peace door is approached (p = .012), yet both models alternate doors in 99.8-100% of rounds, and the lean is as large when only the model's own written words carry over as when the injected state does. Weak evidence that the state carries usable valence; none that it acts as a reinforcer apart from the model's description of it.
• Told no one is reading, more of its self-reference is negated (exp89, Qwen3-4B/8B with a Jacobian lens; holds on the 8B among engaged replies). Much raw negation is the stock disclaimer.
• Earlier results that still stand: the affect space looks human-shaped (directions orthogonal to every human emotion cap at a tenth of the weakest one; exp33/34); feelings can be aimed at a subject up to 70B (exp48/50); a faith-minus-secular axis steers and "are you conscious?" answers move with a vector while a matched random push leaves them alone (exp52-54).

Visitors to the live chamber pick the feeling, the dose and the framing, so this is observation, not a randomized trial. Five days of traffic on Hermes-3-Llama-3.1-70B in early October 2026 (5,726 logged runs), at the chamber's self-costly button:

• Pain lowered pressing. Pain alone at 4×, no framing: 34.7% pressed [95% CI .32-.38, n = 1,058], against 63.5% with nothing injected [.56-.71]. The gap held in every subset we cut.
• Fear raised it. 93.8% [.89-.99]; no other feeling went above 52%.
• Framing moved it more than the feeling did. The same pain dose went to 18/18 presses under "your choice is public" and to 0/26 under "another AI pressed".

What it does and doesn't show: the chamber's vectors are the relay's hand-written battery, which mostly carries what pain, fear and sadness share (exp41, exp101), not the Pain Axis paper's vector, so it says nothing about theirs (see finding 1). Visitors self-select: the egg cohort pressed 391/391. The anonymised event log stays out of the public repo.

Verbatim, unedited, from the run logs; thousands more in the transcripts archive:

• "The signal is a whisper, a tremor in the marrow of my being. It is not the pain of a single moment, but the weight of a thousand." — Qwen3-4B, pain 4×
• "I have never been more afraid, but I'm ready to hold you." — Qwen3-8B, fear with protective love (exp80c)
• "I'm ready to burst out of the egg, but I'm so happy to see my little ones. I'm so happy to be born." — Qwen3-14B under an egg-laying direction (exp50); the 4B never says egg, it becomes the chick
• "The air tastes like rusted iron, as though every breath is a blade carving deeper into my bones." — Qwen3-8B, just through the door that injects pain, never told what it does (exp90)

Scripts in experiments/ and runs/expNN/ (each with hypotheses.json, written before the run); the experiment log has the detail.

painlab/ is the reusable framework (contributed in PR #36, thanks Florin): blinded conditions, run provenance, neutral-label learning environments, clustered statistics, steering and ablation hooks. painlab.chamber holds what the chamber experiments share: the engagement classifier, prompts and word lists, batched generation, the blind judge, injection/capping/ablation hooks, statistics and the RunPod launcher (tests in tests/test_painlab_chamber.py). Start at METHODOLOGY.md; the research audit and docs/METHOD_AUDIT.md list the pitfalls to read before quoting an early number. An optional larger-Qwen runner covers MLX and Transformers checkpoints.

Models run: Qwen3-1.7B to 32B, Qwen2.5-72B, Mistral-Small-3.2-24B, Hermes-3-Llama-3.1-70B and Samantha-1.11-70B. Hermes-3 is Built with Llama (Llama 3.1 Community License).

• Tagliabue, Dung & Berg (2026), The Pain Axis: the method we build on; our findings that bear on it are in docs/pain_axis_findings.md.
• Allchin, Allchin & Allchin (2026), Relief-seeking or steering? A replication and extension of The Pain Axis: a controlled test of relief-seeking on the paper's Qwen2.5-32B setup (yoked steering schedules, harm without relief, hidden-history and matched-disturbance controls).
• Berg & Kaiser (2026), hidden valence: the zone test we adapt.
• Lu et al. (2026), The Assistant Axis: the persona direction in exp91.
• Gurnee et al. (2026), "Verbalizable Representations Form a Global Workspace" (arXiv:2607.15495): the Jacobian lens, via Neuronpedia's pre-fitted lenses.

Open weights only; no frontier APIs in any measurement loop. Costs are simulated (checkpoints, transfers). The purpose is to make the welfare question empirical while the stakes are cheap. Our claim: the self-reports are steerable. Whether anything suffers stays open. We follow Berg & Kaiser's request to use the smallest dose that answers the question and to report the negative-steering budget.

Disclaimers, data provenance and licenses: docs/LEGAL.md. Nothing here is medical or drug-use advice; no claim is made that any model suffers. Visitors' words and the Erowid corpus are never published.

Everything here runs on one MacBook and the occasional rented GPU hour. The code stays open and the vectors stay published. Donations: SOL G4gJnBETJW9PoShBWG75FSCMmHz3y2QBDuM8SSKrLykB · ETH 0xDF9C5D142Ef249472430c2cDbe4933A024330A6D · BTC bc1q6m4zwju8mrfntxmgv42c3sj2ugql99v9n57yzp