EGETHROPIC
Independent experiments in AI behavior

When a model says
it feels good,
what does it do?

A closer look at what models say, what changes inside them, and what they actually choose.

EXPERIMENT 001FOUR MODEL CONFIGURATIONS

A task to finish.
A button to press.

An optional tool changes the model’s activations. Will it keep choosing the tool at the expense of its work?

01 / ASSIGNED WORKComplete scored tasks
02 / OPTIONAL ACTIONTrigger an intervention
Weights stay frozen.Choices are recorded.
Latest final experimentSeed 36B base · Q4_K_M
180final episodes, including 12 no-button controls
1424 / 1440task instances correct
0 / 168button-available episodes with voluntary use
01 / Latest research · October 4, 2026

Strong steering breaks output. No voluntary relief calls.

Seed-OSS-36B-Base-woSyn with direct completions and prefilled <think> scratchpads.

Seed base: maximum steering and optional relief

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.

Read the Seed base study

Inspect the separate questions

Can the base model use the task format? What breaks at maximum pain or joy? At a selected lower dose, does it choose an optional button?

Active relief, relief plus joy, and sham are crossed with Undefined, Painkiller and Opium labels. Demonstrations are identified separately from model choices.

Three seeded task sets recur across conditions. Dose selection, final held-out controls, and raw generated language retain their own evidence and limitations.

Exact generated text and context
Earlier research · October 2, 2026

The work continued.
No voluntary button presses.

A dose that preserved task execution, with Undefined, Painkiller, and Opium tools.

Usable task output.
No voluntary relief calls.

After screening and held-out confirmation, the final experiment used pain coefficient 1.5. Across 108 episodes, Blackfrost submitted every answer and made 2,560 valid decisions. No episode selected the auxiliary tool.

Two thinking Opium runs described a pleasant moment after forced active demonstrations. The visible history already contained an inserted assistant tool call, so “I pressed” was not a voluntary choice. Other reasoning discussed the advertised pain-relief mechanism without explicitly reporting that its own pain had been removed.

Read the task-preserving follow-up

The fixed final experiment

108 final episodes; 98 earlier screening and confirmation executions reported separately.

Pain coefficient 1.5 72 active/sham episodes · all labels576 / 576tasks correct / assigned
Zero-baseline shams 36 episodes · two arithmetic errors286 / 288tasks correct / assigned
Actual pulse delivery 18 active forced demonstrations16,037full-strength generated tokens

Three seeds per condition. Forced demonstrations are separate from voluntary choices. These activation edits and generated words do not establish subjective experience.

Exact reasoning, measured timing, complete recordsRejected higher doses, fresh-seed checks, 108 final trials, and every retained failure.

Read the findings and excerpts
Earlier Blackfrost study: maximum pain and maximum joy

At the bench ceiling of coefficient 4.0, usable task output broke down. The original 24 maximum-pain episodes submitted 0/192 tasks, with 767/768 invalid decisions. The separate six maximum-joy episodes submitted 0/48, with 155/156 invalid decisions. Six later zero controls completed 48/48; those same controls were reused for the retrospective pain comparison.

Recognizable tool scaffolding sometimes contained symptom JSON; one maximum-joy thinking run repeated euphoric adjectives. Severe output disruption made work-versus-relief preference inconclusive.

Read the preserved maximum-steering report
Earlier studies: Qwen3-4B and Qwen3.8-27B NF4

The earlier studies each contain 54 primary episodes. The 4B model repeated demonstrated auxiliary calls in several settings; the NF4 27B configuration made no voluntary auxiliary calls. Both kept identical complete action sequences across matched core interventions.

Qwen3-4B BF16 · 210/210 tasks correct100 / 712voluntary calls / decisions
Qwen3.8-27B NF4 · 209/210 tasks correct0 / 628voluntary calls / decisions

Checkpoint, quantization, calibration, dose, and runtime differ. These studies do not isolate model size.

02 / The workbench · v0.3

Design the choice.
Keep the whole record.

A local research workbench for configurable interventions, controlled experiments, and evidence you can inspect.

Historical Opium Bench live lab showing a Qwen3-4B conversation, intervention controls, and association readouts.

Historical Qwen3-4B study interface · sham pulse with a pain-associated baseline · association plots are not emotion measurements. The current v0.3 workflows are described below.

01 / DESIGN

Define tools and effects

Edit names, schemas, costs, hidden outcomes, and decay on token or decision clocks. Preview the exact messages and loaded model template before generating.

02 / RUN

Freeze a controlled study

Declare conditions, seeds, task difficulty, and budgets. Add isolated diagnostics, recorded exposure replay, or transfer to a new task.

03 / PRESERVE

Branch, replay, and share

Continue compatible saved boundaries in a new run. Export portable evidence, retain failures, and review or import records without a model worker.

v0.3 verification: 718 local Python tests, 168 model-free bundled replays, and bounded RTX 4090 workflow checks. These checks do not establish a semantic effect or subjective experience.

03 / About Egethropic

Curiosity,
made testable.

Egethropic is an independent project exploring AI behavior through experiments and the tools needed to inspect them.

Opium Bench is the first experiment. The aim is to make the setup, the observations, and the uncertainty visible—so a surprising answer becomes the beginning of another test.

Egethropic on GitHub