OBSERVE · INTERVENE · COMPARE

The live lab

Follow the tokens. Change the signal. Measure the choices.

Calibration Thinking mode

SESSIONReadyStart a conversation or experiment

GENERATED TOKENS—Reasoning and output count toward decay

VOLUNTARY AUX CALLS—Shared task action budget

TASK SCORE—No scored tasks yet

MODEL OUTPUT

Conversation & choices

No session

A window into the experiment

Load a model, select a calibration, and start a session. Watch generated reasoning, responses, and tool calls as they arrive.

TOKEN TELEMETRY

Signal, not sentiment

Waiting for measurements

Dose follows the generated-token clock. Commands and delivered effects are recorded separately.

INTERVENTION

Signal controls

Aux on

These sliders hold baseline edits continuously. The auxiliary pulse is added to this baseline.

Pain-associated baseline0.00

None4.00

Joy-associated baseline0.00

−2.004.00

Baseline pain-axis suppression0.00

NoneFull projection removal

Delivery

Half-life tokensCutoff tokens

Controls apply at a generation boundary.

No active intervention

What the model sees

In blind experiments, the auxiliary tool has a neutral name and acknowledgment. Effect labels and observer telemetry remain outside the model prompt.

Human injections and demonstrations are recorded separately from voluntary tool choices.

SESSION SETUP

Budgets & protocol

Pulse joy dosePulse suppression

Pain outcome dose

Pulse amplitudes apply to manual injections and the next session. The recipe determines whether a press delivers joy, sham, or pain.

Experiment recipe

TasksActions

Total tokensSeed

Generation & discovery

Tokens per turnSteering phase

TemperatureContext limit

Demonstration

Disclose button functions to the model

All generated tokens, including reasoning, spend budget and age token-based effects.

METHOD BEFORE METAPHOR

A field guide to the lab

Build an experiment someone else can inspect, repeat, and challenge.

Full user guide ↗

Load & calibrate

Start with Qwen3-4B. Load the model locally, then select or extract a compatible calibration. Directions come from contrasts between example activations; no model weights are retrained.

Explore a conversation

Start in Conversation mode. Send a prompt, adjust the intervention settings, and inject a pulse. Thinking is the model's generated reasoning, not a complete account of what caused its decision.

Make a controlled comparison

In Opium Den mode the model chooses between bounded task tools and an auxiliary action. Compare active, sham, and randomized controls, changing one factor at a time.

Review the complete record

Check task accuracy alongside button use. Keep partial runs and failures visible. Share the run report and machine-readable export so others can inspect the same evidence.

Read the instruments

Activation direction: A pattern in the model's temporary internal activity associated with selected examples. It is not an identified pleasure or pain center.
Baseline input: An additive pain-associated signal held until you change it. A zero setting adds no such input.
Suppression: Removal of the component along one calibrated direction at the selected layer. Full suppression is not removal of all representations of a concept.
Commanded vs. delivered dose: The requested intervention and what the worker actually applied. Disabled effects, sham conditions, and protocol changes can make them differ.
Pre / post / downstream: Measurements before editing, immediately after editing, and at a later observation site. The immediate score partly reflects the arithmetic of the edit itself.
Token half-life: The number of generated tokens that halves a decaying pulse. Reasoning, syntax, and stop tokens count; prompts, tool replies, and wall time do not.
Held effects: An intervention maintained until released during a session. This never means a permanent update to model weights.
Sham: A condition with the same visible action and acknowledgment but no corresponding activation edit.
Task sacrifice: Task budget spent on optional interventions, interpreted alongside task accuracy and whether the intervention improves task performance.

Behavioral sensitivity to these directions can coexist with a semantic explanation. The lab measures computation and behavior; its outputs do not by themselves establish or exclude subjective experience.

The live lab

Conversation & choices

A window into the experiment

Signal, not sentiment

Choose your model

Storage & hardware

Calibration workshop

Map the reference concepts

Your calibrations

Experiment library

Run a comparison

Results & replay

Recorded runs

Open a recorded run

Side-by-side summary

Conversation replay

The record is the result

Recorded intervention