A small creature driven by a real neural network, a deep Q-network (DQN), that trains in this tab from
what happens to it. Nothing is sent to a server. Nothing trains while the tab is closed or hidden.
The brain
DQN (default): the network picks every move. Every few seconds it replays recent experience and takes a training step (a "lesson"). Some moves are deliberate coin flips so it keeps exploring: 1 in 10, the same for every creature in the family, all its life.
Rules: a fixed, hand-written agent moves the body. Use it as a baseline. The DQN keeps training off-policy from what it sees.
Keeping one: when your familiar dies, the world keeps going and you keep one of its family. The learning copy and your lessons go on with the one you keep; the kept creature's own birth brain acts, and its bond starts again.
Rest: left alone for five minutes, the whole family rests: nobody eats, lays, ages or dies. Your familiar studies its own memories, the ones it got most wrong first; the arcs on the world are the memories it replayed. It stops early if its error on a held-back set of memories starts rising. Any touch wakes everyone.
The family
A family: up to 12 creatures, eggs included, each with its own small neural network (its birth brain) and a body drawn from its genes. They eat, lay eggs paid for with food they ate, and die of hunger, bites or old age. Only your familiar learns; the others act on the brains they were born with.
Heredity: an egg is a copy of one parent's genes with small random changes, so a child looks like its parent. Genes change only at birth. Every gene is drawn on the body, and a creature's card (Family) says which the world reads and which are look only.
Instincts: labelled wherever they act. A newborn trails its parent for its first seconds out of the egg; a wary creature keeps clear of live hazards; a hungry creature saves its strength, resting instead of wandering, exploring, chirping or coming to you (hunting for food is still its brain's choice); a creature lays only when it is grown, has eaten enough and can keep some energy back.
Meeting one: tap a creature to meet it, tap empty ground to bring a meal. Follow moves only the view; the Follow pill then names your familiar and brings the view back to it.
Kinds (Family): a kind is a body and a diet. It counts once a parent of that kind raises a child of the same kind to adulthood.
Seasons: only when you start one, a copy of the family lives 10 to 20 generations in the practice world, in this tab, while home keeps living. Then its descendants race their revived founders and a luck-only line in new worlds, and the page says whether they did better than luck there. Up to 2 eggs can come home to the nest.
The world and the view
One map: a 2400 × 2400 world, the same on every screen, with 13 named places that stay put every life: 5 groves, 4 springs and 4 thickets. The screen shows only part of it, and the creature senses only what lies within about 1,000 px of it.
Looking around: drag the world to look; pinch (or scroll) to zoom from 0.5× to 2×. Left alone for 8 seconds, the view comes back to the creature. On a keyboard, the arrow keys look around while the world has focus, + and − zoom.
The map and Follow: the map in the corner shows the whole world; tap a spot to look there. Follow (or the 0 key) brings the view back to the creature; pressed again, it also resets the zoom to 1×. While it is out of view, an arrow at the edge points to it; tap the arrow to follow.
Measuring it
Proving Ground (Mind): 32 fixed courses in the practice world its ancestor grew up in, scored after lessons 10, 25, 50, 100 and every 100 after: food − 2 × bites − 25 if it dies, beside its ancestor and a random mover. It can go down; "within noise" means the change is smaller than twice its standard error.
Champion and learner: the body acts on a champion (at first, the ancestor) while the learning copy trains. A learning copy takes over only when it measures better beyond noise on those courses and again on 32 held-out ones. The Brain vault (Tune) lists the copies; rolling back, restarting or promoting by hand is always your click, and says so.
Lessons (Tune): pick a drill, show it moves (keys 1–7), then re-test on the same runs. Most short lessons read "no measurable change", and the card says so.
Lessons you gave (Mind): after a paid Praise, Scold or Feed, how much the learner's value for that moment and move has changed since.
Teaching it
Feed: hands your familiar a meal from the nearest patch of land that has one, and it eats it on the spot: the energy and the egg store are any meal's, and no one else can take it. It also counts as a reward for what it was doing. While the energy bar is red, Feed helps. Tapping (or clicking) empty ground drops a meal where you tap, and anyone may eat that one; a drag only looks around. When the land is bare, nothing comes.
Praise: a positive reward for whatever it just did.
Scold: a negative reward for whatever it just did. It is a lesson, not a punishment: the move is stored with the penalty and replayed in training.
Challenge: a pulser near your familiar, a small setback to learn from. It bites anyone nearby, not just your familiar.
Reward tuning: the sliders scale each part of the reward, so they change what it learns to want.
Saving and starting over
Save Brain: stores the network's weights in this browser (IndexedDB). It also saves on its own as it trains.
Reset Brain: replaces the learning network with its ancestor's weights; the ancestor drives the body again. The vault keeps its saved copies to roll back to.
Clear Memory: starts a new family in a new world. It clears the family, the tree, the nest, the kinds, the seasons and the loophole log, and the letters and keepsakes. It keeps the brain vault, the Proving Ground history, the lesson book and your lessons.
Brain file (Tune, vault): download the acting weights with their Proving Ground proof, or load one: it is re-run here before it can act.
Chat and voice
Chat: replies are written in this tab from the creature's live state. Chat never decides its moves.
Language-model voice (optional): a small model downloaded once, after you agree, can rephrase those replies. A rephrasing that changes a verdict or adds a number, a name, a feeling, a question or a claim the reply does not make is dropped for the built-in reply.
Read replies aloud: uses your browser's speech synthesis. Off by default.
Soundscape (Tune): a quiet drone that follows its mood. Off until you switch it on; silent whenever this tab is hidden.
Activity log: everything it said and did, with filters.