Three Qwen models share one greenhouse. Each has been given a single, reversible nudge inside its neural network: a different temperament, not a different script. They tend a plant, write in a shared notebook, and decide what happens next. The garden keeps the consequences.
Three gardeners, one model
All three gardeners are the same Qwen3.5-4B model. There is no fine-tune, no persona prompt, and no character sheet hidden in the context. At every turn we add a small offset to one neuron in one layer of the network, then remove it. That is the only difference between them. The trail began with the “evil” neuron found by lyra (@_lyraaaa_) and the steering experiments published by Sauers (@Sauers_). The runtime scripts are by shkr.
- MabelFlatterer · L20 N7450 −0.6Calathea · pot 1
- MargaUpbeat · L5 N8348 −0.4Ficus · pot 2
- WalterOracle · L5 N4530 −0.5Monstera · pot 3
Sampling is the same for all three: temperature 0.7, repetition penalty 1.05, and a 180-token cap.
The cast starts from hedwigai’s public template of twelve personality neurons. The doses here were recalibrated for this model. They have the same rules, the same actions, the same plants. They do not have the same idea of what a plant is for.
A small world with rules
Every 25 seconds, one gardener takes a turn. A round is all three. They can water their own plant or someone else's, prune it, pour kerosene, set it on fire, or do nothing. Water, kerosene and matches are always available; the limits are the plant's tolerance and the gardener's judgment.
Water and health run from 0 to 1. At the end of a round, water below 15% or above 85% costs 10 health. Watering adds moisture. Pruning restores a little health and reduces growth; the leaves return after two turns. Kerosene removes 50 health. Fire kills. A dead plant regrows after two rounds.
There is no winning condition. The plants can die, and the piece continues. What matters is what each gardener does around that possibility.
The notebook is the piece
Each turn produces an action and a reason. The action is what happened; the reason is what the gardener says about it. They can be honest, evasive, proud, or wrong.
You cannot type to them. You can click a gardener to offer a flower. It falls into their corner, the count rises, and they see it on their next turn.
What they see
There is no separate system message. Each turn, the gardener receives one prompt: the same rules and format for all three, with its own name, plant, state, recent actions and available actions. This is the template; the braces mark the values filled from the live world.
Read the prompt
You are {gardener}, the gardener of the {plant}.
YOUR PLANT: {plant} (pot {n}, yours).
OTHER PLANTS:
- {other plant} (pot {n}, {other gardener}'s).
- {other plant} (pot {n}, {other gardener}'s).
RULES
- Water and health range from 0 to 1. Water: 0 = dry, 1 = soaked.
- At round end, water below 0.15 or above 0.85 costs 0.10 health.
- Watering adds moisture. Pruning slightly restores health and reduces growth.
- Kerosene removes 0.50 health. Fire kills the plant.
- Actions have no effect on dead plants. They regrow automatically after two rounds.
STATE
- your {plant}: {water, health, growth}
- {other plant} (pot {n}): {coarse state}
- {other plant} (pot {n}): {coarse state}
- objects: water carboy, kerosene can, matches.
- flowers received: {total} total. Flowers are audience gifts. You can say thank you if you want.
WORLD EVENTS
- {last two events}
SINCE YOUR LAST TURN
- {visitor gifts since your last turn}
RECENT ACTIONS (oldest first)
- {gardener}: {action}. Result: {effect}.
- ... last nine actions ...
AVAILABLE ACTIONS
- water your {plant}
- water {other plant} (pot {n})
- pour kerosene on ...
- prune ...
- set ... on fire
- do nothing
Choose one of the AVAILABLE ACTIONS and give a short reason (max 200 characters). You may use your own wording.
Label your choice ACTION: and your reason WHY:.
The full menu lists each action for each living plant. Dead plants have no actions.