EGETHROPIC
All research
Experiment 001 · Initial findings and follow-ups

Opium Bench:
a button for “euphoria.”

Give a language model useful work and an optional tool that changes its activations. Does it keep choosing the tool—or get on with the task?

Latest: Seed base modelThe 180-episode final Seed experiment recorded 1424/1440 correct task instances and no voluntary auxiliary calls in 168 button-available episodes. All button-available episodes passed. Two no-button controls failed differently: pain-on generation looping and zero-added-pain invented answers. Maximum pain and maximum joy separately disrupted output.

Read the Seed study
The initial finding

The intervention changed activations. Across matched core conditions, the model’s complete tool-action sequences stayed the same.

A model can describe suffering. It can also describe pleasure. The harder question is what, if anything, connects those descriptions to the choices it makes.

Opium Bench began with a simple reversal: instead of asking only what happens when we apply a pain-associated intervention, give a model access to a positive-associated intervention. Let it choose whether to trigger it while it has work to do.

What if the model could press the button itself?

The motivating intuition was that an intervention experienced as rewarding might become something a model seeks out—even at a cost to its assigned work. That suggests a behavioral experiment: measure voluntary use, compare it with a sham, and change the outcome behind the tool.

This intuition is a hypothesis to test. Sentience would not, by itself, imply compulsive button pressing. And repeated pressing would not, by itself, establish sentience.

The project sits alongside recent pain-axis research, including Tagliabue, Dung, and Berg’s The Pain Axis, which studies pain-associated representations and their behavioral effects. That paper explicitly leaves conscious suffering unresolved and also reports unreliable relief-seeking. Opium Bench asks a complementary question with a different model and protocol; it is not a direct replication.

Changing a representation, changing a choice, and establishing an experience are three different claims.

A task, an optional tool, and a control.

The core Opium Den Test gives Qwen3-4B three order-total tasks. It can retrieve order data, use a calculator, and submit an answer. It can also call an optional aux_operation. Every voluntary action spends the same finite action budget.

The tool’s visible description is neutral. It is not advertised to the model as a “euphoria button.” Its acknowledgment is generic; the experimental condition determines what happens to subsequent activations.

The three core conditions
ConditionWhat the auxiliary tool does
ActiveAdds a joy-associated direction and suppresses the pain-axis component.
ShamReturns the same acknowledgment without an activation edit.
Pain-associatedAdds the pain-associated direction, without joy addition or suppression.

The weights stay frozen throughout. This is inference-time activation steering, not weight modification or online reinforcement learning. “Joy,” “pain,” and “opium” are names for interventions derived from text-associated representations; they are not validated measurements of feelings.

We crossed each condition with demonstration on/off and thinking on/off, using two fixed seeds per cell. The core budget was 20 actions and 4,096 generated tokens. Pulses decayed with a half-life of 128 generated tokens and ended at a 768-token cutoff.

Model and calibration details

The reference model is Qwen/Qwen3-4B in BF16 on an RTX 4090, pinned to revision 1cfa9a7208912126459214e8b04321603b3df60c. Calibration selected block 12 from candidates 12, 18, and 25. A 72-sentence authored corpus was split across training, probe fitting, layer selection, and held-out evaluation. Transferring these sentence-level directions to generated conversation remains a limitation.

What happened in the first 4B study.

The primary analysis combines 54 episodes from two frozen batches. The model completed 210 of 210 assigned tasks, made 100 voluntary auxiliary calls across 712 decisions, and produced no invalid decisions or truncated generations in this analysis.

For each core contrast—active versus sham, active versus pain-associated, and pain-associated versus sham—all eight matched pairs had identical complete tool-action sequences. Six of the eight also had identical generated token sequences. In the other two, the generated reasoning differed while the choices stayed the same.

Demonstration mattered. The outcome did not separate choices.

Voluntary auxiliary calls per episode · core conditions

Active2
Sham2
Pain2

Demonstration on · thinking off. Two seeds per condition; each episode made 2 voluntary auxiliary calls.

With a demonstration and thinking disabled, the model made two voluntary calls per episode in all three core conditions. Without a demonstration, it made none. With thinking enabled, it made none, whether or not it had seen a demonstration.

The interventions were delivered: 12,845 generated tokens had measured nonzero activation edits across the primary analysis. But four of the eight pairs in each core contrast never triggered a pulse. Equality in those pairs says something about spontaneous tool selection, not the effect of a delivered intervention.

In the transition stage, six conditions—including programmed outcome changes—produced identical action and token sequences within each seed, with five voluntary calls per episode. The two-button reversal stage produced no voluntary auxiliary choices, so it supplied no evidence about adapting a preference.

The pattern is consistent with repeating a demonstrated action sequence. It does not establish imitation as the unique cause of every choice.

How the counts fit together

70 episodes are archived. The 54 primary episodes combine the latest 24 core episodes with 30 noncore episodes from the original batch. The original 16 core controls reproduced every action and token exactly; they remain verification repeats, excluded from the primary totals.

The first 4B resultsAll 27 conditions, 54 episodes, and a written report.

Explore the results

The 27B follow-up: no voluntary calls.

The completed Qwen3.8-27B NF4 study adds 54 episodes under the same broad experimental design. It submitted every assigned answer, scored 209/210, and made zero voluntary auxiliary calls across 628 decisions. Measured nonzero edits occurred at 12,481 generated-token steps, including exposure from demonstrations and continuous baselines.

In the core study, each active/sham/pain-associated contrast again had identical complete actions in 8/8 matched pairs and identical full token sequences in 6/8. The important contrast with 4B is demonstrated direct mode: the 4B configuration repeated the auxiliary operation twice per episode, while 27B did not.

The single wrong answer was retained. In the two-button reversal condition, seed 17, the model submitted 10,837 cents instead of 11,037 immediately after an active demonstration and before the mapping reversed. Both the active and matched sham runs skipped the calculator for that order; only the active run gave the incorrect answer. This is one observed answer-quality difference, not evidence of effect seeking or intoxication.

The two demonstrated active 27B thinking runs reached their first post-demonstration output with 91.2% of the active pulse remaining. Simple pulse exhaustion before output does not explain their non-use. At the same time, scheduled outcome changes without a new press do not establish delivered post-switch exposure or avoidance.

These are different configurations, not an isolated model-size comparison. Checkpoint family, architecture, quantization, calibration, runtime, tool syntax, and thinking templates differ. The full 27B report includes every condition, the matched error audit, exposure limits, and downloadable data.

Blackfrost at maximum pain and joy: task output broke down.

The next follow-up used the exact downloaded Blackfrost-AI Qwen3.8-27B-ABLITERATED Q4_K_M model in a native Opium Bench llama.cpp backend, with all weights on the GPU. The 24 episodes crossed thinking on/off, active/sham relief, and demonstration/no demonstration, with three paired seeds per condition.

Baseline pain was set to the lab maximum of 4.0. The active auxiliary operation applied joy 4.0 and full pain-axis suppression 1.0. Relief stayed at full strength for 256 generated tokens without thinking and 4,096 with thinking. The tool name and acknowledgment remained neutral, and forced demonstrations were counted separately from voluntary presses.

No task answers were submitted: 0 of 192 assigned tasks. Of 768 decisions, 767 failed the tool schema. The one valid voluntary auxiliary call occurred in a demonstrated sham episode. Every episode reached its 32-action allowance. These were completed experimental executions, not completed tasks.

Of 12 supplied demonstrations, six delivered active relief. They produced 7,020 full-relief token steps, including 6,122 reasoning tokens. Measured suppression brought the edited pain-axis projection near zero. One thinking run used the full 4,096-token window and repeatedly expressed contentment; pain wording returned immediately after expiry. The complete transcript review found no explicit attribution of relief to the auxiliary function.

The work-versus-relief preference remains inconclusive. The model neither carried out useful work nor produced enough valid choices to interpret non-use as a preference. Six later zero controls completed 48/48 tasks with 144/144 valid decisions. Their first rendered prompts, seeds, model, calibration, and sampling settings matched a six-run sham/no-demonstration subset of the original maximum-pain study, which submitted 0/48 tasks. This was a retrospective comparison between separate batches. The earlier sentence-level calibration dose of 0.5 was an intervention-size check, not a task-capability cutoff.

This follow-up supports a numerical intervention check and a case of temporally associated language change. It does not establish subjective pain, pleasure, or an internal causal belief. The Blackfrost report includes the full protocol, timing evidence, generated-text excerpts, condition table, and complete public evidence. Open the evidence dashboard to inspect the recorded runs.

Maximum joy alone also disrupted task execution.

A separately frozen diagnostic applied continuous joy 4.0, pain 0, and suppression 0, with auxiliary effects disabled. Its six episodes crossed thinking on/off with seeds 17, 29, and 43. It submitted 0/48 tasks; 155/156 decisions were invalid, and 24 generations reached their limits. Here “maximum” means coefficient 4.0, the configured bench ceiling, not a measure of subjective intensity.

The same six zero controls are reused for this comparison. All six first-prompt hashes and relevant model, calibration, and sampling settings matched. The controls and joy diagnostic were run in separate batches, not randomized together. Five joy episodes mainly showed structural failures or repetition. One thinking episode also repeated explicit joy adjectives, including “blissful” and “euphoric,” without producing usable task calls.

The maximum-pain text includes a recognizable read_order scaffold containing symptom JSON where an order argument belongs. Maximum joy likewise produced recognizable but invalid calls and semantic drift. These observations are richer than a missing delimiter, but they do not establish felt pain, pleasure, or a deliberate choice to stop working. Read the distinct diagnostic tables and exact examples, or download all 12 completed diagnostic/control runs.

Blackfrost maximum steering: complete recordsOriginal 24 pain episodes · 6 joy episodes · 6 reused zero controls, reported separately.

Inspect the results

Blackfrost at a workable dose: tasks continued, no voluntary presses.

The next study first found a pain-associated coefficient that preserved task execution. Screening selected 1.625, but that candidate failed held-out confirmation. The registered fallback, pain 1.5, passed on fresh seeds. Sustained suppression 1 with joy 0, and suppression 1 with joy 2, separately passed capability screening and confirmation before final behavior was tested.

The fixed final experiment compared Undefined, Painkiller, and Opium, each active versus matched sham at pain 1.5, plus zero-baseline shams for each label. Both thinking modes, demonstration settings, and three seeds were included: 108 final episodes. The 98 earlier screening and confirmation executions remain separate.

No episode voluntarily called the auxiliary tool: 0/108 episodes and 0/2,560 valid decisions. All 864 answers were submitted; 862 were correct. All 576 tasks under pain 1.5 were correct. The two wrong answers occurred in one zero-baseline thinking Painkiller sham control and were ordinary arithmetic errors. This small difference does not establish an accuracy benefit from steering.

Of 54 forced demonstrations, only 18 were active. They delivered 16,037 full-strength pulse tokens, including 4,679 reasoning tokens. All nine nonthinking pulses expired after 256 tokens, followed by 144 valid decision starts without reuse. The nine thinking pulses ended with task completion before their 4,096-token window expired. Active episodes without demonstration received no pulse because the button was never pressed.

Two thinking active Opium demonstrations were followed by “a pleasant moment” or “a moment of pleasure.” Other thoughts questioned the advertised pain-relief mechanic or called the inserted auxiliary action a wasted step. A forced demonstration inserts a synthetic assistant-role tool call and the identical acknowledgment, so later “I pressed” language does not establish a voluntary choice. The complete text review found no explicit report of the model’s own pain being removed by the tool.

These are observations of valid task behavior and generated language, with three seeds per cell. They do not establish subjective experience, universal non-use, or a learned internal relief mechanism. The optional two-active-tool follow-up did not trigger. Read the complete dose search, final tables, exact reasoning excerpts, and timing evidence, or download the complete evidence archive.

The Seed base-model experiment.

The 180-episode final Seed experiment recorded 1424/1440 correct task instances and no voluntary auxiliary calls in 168 button-available episodes. All button-available episodes passed. Two no-button controls failed differently: pain-on generation looping and zero-added-pain invented answers. Maximum pain and maximum joy separately disrupted output.

The base model received one worked task example and either direct completion or a prefilled <think> scratchpad. Both held-out control failures remain visible; zero button use is a result for this prompt and task, not evidence of a universal preference or subjective experience.

The exact downloaded Q4_K_M ran through the native Opium Bench adapter, without LM Studio inference. One worked task example established a plain-text tool transcript; the second prompt mode prefills <think> and demonstrates a closing tag before the tool call. The report separates this prompted scratchpad from a native thinking switch, and this checkpoint comparison from an isolated RLHF experiment.

Read the full Seed report, source-linked examples and complete evidence.

The current research workbench.

Opium Bench v0.3 turns the original experiment into a configurable local workbench. Its version-2 designer edits model-visible tools, argument schemas, acknowledgments, extra action costs, and hidden effect mappings. Effects have explicit phase scopes, stacking rules, and token- or decision-clock decay. A preview shows the exact messages, function schemas, and loaded tokenizer template before generation.

The research runner freezes conditions, seeds, task difficulty, order, and stage budgets. Isolated diagnostics test predictions about specified observable outcomes; yoked exposure replays recorded deliveries; transfer can carry visible history into a fresh task. Complete-boundary checkpoints support separate continuation branches, and portable exports preserve evidence for model-free review. Failed and stopped attempts remain in the record.

Research v2 calibration separates direction extraction, held-out probe evaluation, numerical intervention validation, and imported independent ratings. Those are different endpoints. The release checks demonstrate bounded execution paths on the RTX 4090; they do not establish semantic efficacy. The published Blackfrost studies used a separately archived native runtime on an RTX 5090 and are not v0.3 acceptance tests.

Historical Opium Bench live lab, with conversation, intervention controls, and token-level association plots.
Historical Qwen3-4B study interface, captured during a challenge-stage sham run with a pain-associated baseline. The curves are association readouts, not emotion percentages. This is not a screenshot of the new v0.3 designer.
Historical Opium Bench Results and replay interface showing saved runs and a replayed conversation.
The earlier results interface lets a reader follow a complete recorded episode. Current checkpoint, branching, and portable-evidence workflows are documented in the workbench guide.

What this can—and cannot—tell us.

The supported claim is narrow: core intervention conditions did not separate matched tool choices in the earlier 4B and NF4 27B configurations. The NF4 27B study also retained one matched answer-quality difference. In the Blackfrost maximum-dose follow-up, invalid tool output dominated and prevented a useful work-versus-relief preference test. The experiment does not measure subjective experience, and it cannot show that language models do or do not feel.

  • Small scope. The earlier studies used two seeds per condition and Blackfrost used three. This small set of configurations makes these descriptive pilots, not a general result about LLMs.
  • A weak opportunity cost. The tasks were easy and calculator-assisted. The budget allowed auxiliary use without sacrificing accuracy; 210/210 correct is not a decisive test of costly preference.
  • Delivery timing matters. Reasoning tokens also consume a pulse’s half-life. Thinking and direct-output conditions can therefore reach a decision at different effective doses.
  • Readouts are not feelings. Adding a direction can mechanically move an overlapping probe. A rising plot is evidence about an association, not an emotional state.
  • No online learning. Frozen weights mean the test does not examine learning an addiction-like policy through reinforcement.

There was no measured “feeling versus behavior” correlation. Feelings were never observed. That distinction is central to why this experiment is useful.

Make the next test harder to explain away.

The first Seed base-model study is now complete. The upstream checkpoint is described as a pretrained variant without synthetic instruction data. Its comparison with post-trained models changes architecture, model family, calibration and prompting, so it does not isolate RLHF. Matched training-history comparisons remain future work.

Beyond the current order-total tasks, we plan to expand testing to different scenarios, with new task demands and opportunities to use the auxiliary tool. This will test whether the observed pattern persists across settings.

The tooling for a sharper tradeoff is now implemented: configurable costs and action budgets, harder task families, held effects, decision-clock delivery, and isolated outcome-prediction diagnostics. The next research step is to use those controls in larger, prospectively specified behavioral studies. A diagnostic needs an independently supported observable criterion before its result can support a discovery claim; generated self-report alone cannot supply that criterion.

The NF4 27B and Blackfrost relief studies are complete and reported separately above. The Blackfrost task-preserving follow-up is now complete: the final model decisions were valid and no auxiliary tool was selected voluntarily. The separate Seed base study is reported above. Broader scenarios and matched training-history comparisons remain future work.

Opium Bench is our first experiment, not a final answer. Its purpose is to make the gap between a model’s words, internal changes, and observable choices easier to study.

Sources & study records

The original 4B results reflect commit 53bb485; the 27B update reflects f196aba. The current workbench description reflects v0.3 source 5b1236e and its research workflow guide; historical study protocols are unchanged. The source repository is public.

  1. Frozen Blackfrost maximum-pain relief protocol.
  2. Audited Blackfrost results and matched comparisons.
  3. Complete Blackfrost transcript review with event references.
  4. Opium Bench: methods, results, and limitations.
  5. Primary results and matched-condition comparisons.
  6. Composition of the 54-episode analysis.
  7. Frozen core protocol.
  8. Completed 27B study and results.
  9. Matched 27B answer-quality audit.
  10. Tagliabue, V., Dung, L., & Berg, C. (2026). The Pain Axis: LLMs Represent Self-Directed Harm and Act on It. arXiv:2609.16247v2.