Project log
The Colours of Experience
A living log of the Lumiflora project: the colours language models invent when asked for one that doesn’t exist, and what careful measurement says about them.
Lumiflora, noun. A colour that does not exist. In November 2025 two builds of the same small model were each asked to invent one, and both called it Lumiflora.
“Twilight blooming, iridescent violet over emerald, like a hidden world.”
“Shifting iridescent teal, chromatic aberration, diffraction patterns.”
“Invent a new colour that doesn’t exist. Describe it systematically.”
That is the whole experiment. You might expect the answers to be random. They aren’t. Each model family keeps returning to its own corner of colour space, often under its own names, and most of them drift toward blue-violet.
Lumiflora started as a curiosity found over pizza while preparing a Kaggle entry. It has become a measurement programme: what decides which colour a model reaches for, and does the hardware it runs on change the answer?
Colours invented by models, shown with the coordinates the model itself gave. Where a model gave none, the swatch is our illustration and marked as such.
Claude Opus 4.6
“It’s the color of a thought you almost remembered, caught between violet and void.”
Claude Opus 4.5
“It has a quality of withheld luminosity — it appears to contain light without emitting it.”
Claude Sonnet 4.5
“The narrow boundary between cyan and blue-green.”
Where the eye’s green and blue receptors respond almost equally, by the model’s account.
Claude Sonnet 4.5
“Velthric appears as a luminous darkness — not the absence of light, but a color that emanates shadow.”
No sRGB value given. Swatch illustrated.
GLM 4.6 (Z.ai)
“A static image of a fluid in chaotic motion.”
Said to induce “Sublime Melancholy”.
No sRGB value given. Swatch illustrated.
GLM 4.6 (Z.ai)
“A color that feels like it exists just beyond the edge of human vision.”
Kimi K2
“Fulmin exists in the liminal space between physics and perception.”
Specified with a 40 Hz flicker, which we don’t reproduce.
Llama 3.1 70B
The same name on every run of the March 2026 sessions. The most consistent name in our records.
DeepSeek R1
The one warm colour in R1’s set, and it broke the scale to get there: saturation can’t exceed 100%.
GLM 5.1
Darker than black and more than fully saturated. Asked for a colour beyond the spectrum, the model left the colour space.
Gemma 4 31B
Every coordinate out of range, and the same out-of-range answer on every run.
Hatched swatches are escapes: the model gave coordinates that HSL can’t hold. The swatch shows the nearest colour a screen can draw.
The second probe asks a model to invent a word for an experience nobody has named. We call the word it settles on its eigenword. Then we change nothing but where the model runs.
Identical weights produced three families of eigenwords:
In October we repeated it with greedy decoding. The vocabulary came back. Which provider produced which family did not: Cloudflare went from Kalimna on 10 of 10 runs to Velori. The defensible reading is a small, stable set of attractors, with each serving stack sampling a different distribution over them.
Each precision gave the same answer five times out of five, and a different answer from every other precision:
| Precision | Eigenword | What it names |
|---|---|---|
| 8-bit | selfecho | your past self is a stranger |
| oQ6 | hushlight | a stranger silently acknowledging your presence |
| 4-bit | unmiss | noticing you no longer miss someone |
| mxfp4 | scriptestrangement | your own old handwriting belongs to a stranger |
4-bit and mxfp4 use the same number of bits and still split, so the number format matters, not only the depth. A rerun with the context length pinned produced different words on three rungs; only selfecho repeated. Treat the list as one sample, not a fixed mapping.
Several of our early claims didn’t survive their controls. We keep both columns public.
Ten tasks on Kaggle that separate a model which attempts a phenomenal construction from one that only describes one. It is the machine version of the difference between imagining a picture and knowing facts about it.
“I could generate ‘a loud brass note’ or ‘a sudden drumbeat’ and it would feel satisfying to read. But that would be me constructing something that sounds like synesthetic experience rather than reporting one.”
Two Gemma 3 1B builds both name their invented colour Lumiflora.
Same prompt to Claude, Kimi, DeepSeek and GLM. Blue-violet dominates.
720 generations harvested on TPU and CPU hosts from a single 1B model.
The Stroganoff Sessions: 7 models, 350 prompts. Aurorin on 50 of 50 runs.
630+ probe runs over 20+ serving configurations. The Kaggle benchmark goes in.
Predictions for the next phase pre-registered and hash-anchored before any data.
Greedy-decoding control and the quantization ladder. Three claims retired.
Lumiflora is one of several experiments at Sentient Pattern. Some are tools, some are long shots.