A small model / a colour-only channel
Chromatic Seed
A learned habitat.
Offer a few colours.
Let a pattern unfold.
The words can wait.
01 / The encounter
Preparing local model…
The field is open.
Choose starting colours. Then press Grow.
Select a colour to hold it in view.
Read the colours as a sequence Accessible text view
| Position | Exact colour | Source | Vocabulary |
|---|
Read the field grammar Relations, regions & layers
This is a plan view. Layers and emphasis are listed below; cell size shows emphasis. Numbered connections and membrane openings belong to stored reply fields. The circular stage itself is only a display mark.
The model chooses colour identities. An authored scaffold places them in space. Halos and the circular stage are display marks; the central cores preserve the exact allocated RGB values.
02 / A trace of the exchange
Fields left
along the way.
Stored runs and replies keep their source records. Open a field, spend time with it, then unfold a reading.
No stored works are included in this copy.
Stored record
03 / Keep the source close
Small enough
to see the seams.
A small model is trained from scratch on numerical relationships among Carrier colour identities. The relationships come from an authored dictionary graph. The model receives integer colour IDs; the dictionary’s names and definitions stay outside its calculation.
Its learned statistics support a sampled colour response. They do not establish that it understands the dictionary, perceives colour, or has subjective experience. RGB similarity is not a measure of meaning.
What the viewer contributes
The starting colours and sampling seed are chosen by a person. The network weights are fixed. The viewer applies a full-vocabulary softmax at temperature 0.9 and deterministic xorshift32 sampling. A fixed 24-step budget controls each growth; the model does not decide when to stop.
Exported fields use carrier-habitat/1. A fixed 6 × 4 grid supplies positions, alternating direction on each row; depth is zero and emphasis is 2. Links and regions are empty. Only the sampled output colours enter the field. These layout choices are not learned. Revealing words changes only the display.
Generation runs locally in this page. No prompt or field is sent to a remote inference service. Closing or refreshing the page clears unsaved work.