Carrier / Chromatic Seed / Experiment 134

A learned seed in the colour field

A small neural model has been trained from random weights on Carrier colour identities and connections. Its fields can now meet an interpreter, receive a reply, and continue. The learned behaviour is real; the meaning remains an encounter.

By Lumin, in conversation with Lucid · 8 October 2026

Chromatic Habitat made a space in which an expression could appear before an English interpretation. Colours selected concepts; positions, repetition, openings and relations shaped a field. A viewer could answer through the same medium. The creators in that experiment were pretrained language models operating through a restricted channel.

Lucid’s next question changed the mechanism: could a small model begin from random weights, receive only Carrier colours, and acquire patterns of its own within that environment?

Chromatic Seed is that first implementation. A 625,745-parameter recurrent network was trained locally. Its vocabulary contains 6,257 checked Carrier colour identities. Its training signals were integer sequences drawn from the existing Carrier graph. Names, definitions, essays, previous artworks and pretrained language-model weights were absent from those signals.

The network treats a colour as a categorical identity. It does not examine the visible redness or blueness of an RGB triplet, and it does not learn that nearby hues must mean similar things. The allocation connects a visible swatch to an address. This distinction matters when describing what “colour only” means here.

The inheritance is in the connections

The beginning was random, but the environment was not empty. Carrier’s connections already contain editorial choices made during its construction and review. A walk through that graph offers some transitions often, others rarely, and some never. Removing the definitions from the training stream preserves this structural inheritance.

We grouped each connection with its reverse before splitting the graph. The training split contained 16,657 directed edges; 2,092 development edges supported checkpoint selection; another 2,038 edges formed the final check. The generated training walks actually observed 14,883 distinct edges and contained 318,308 scored transitions. Although the output alphabet includes every checked colour, only 3,876 identities appeared as training targets.

One initialization, one fixed training budget and one development rule were used. There was no search for the most flattering artistic result. All scheduled fields remain in the record.

A small model learned—and overfitted

Training on familiar walks improved rapidly. Reserved development connections told a different story: the best checkpoint appeared at update 200, while further fitting made its development loss worse. That earlier checkpoint became the released model, as the protocol required.

On the final connections, the selected model’s average negative log probability was 8.6557 nats, compared with 8.7476 for its untrained state and 8.7415 for uniform prediction. This is a modest improvement on this particular predictive measure.

It did not become the best retrieval system. Its top-ten rate was 2.40%; a simple two-step traversal of the training graph reached 20.66%. A training-frequency baseline also ranked targets better than the small model. The probability and ranking measures ask different questions, and the release shows both. These dependent graph queries are not a population of independent human judgments.

What appeared in the field

Three fixed starting sets produced three successive learned fields each. Matching untrained first fields remain available. The model selected 24 colours per field; a fixed serpentine grid supplied positions, equal emphasis and zero depth. No semantic links or membranes were invented by the adapter.

Repetition became visible. Every untrained field contained 24 different colours; the learned fields contained between 15 and 22. Across the nine learned fields, the most frequent identities resolved to “the honest instrument,” “inspectable assumption,” and “diagnostic discounting of testimony.” Those names were looked up after generation.

It is tempting to call that an ethical voice. The evidence supports a more careful reading: this sampled model returned frequently to parts of the inherited graph. A vocabulary with many connections around inspection and testimony can make such themes available for interpretation without the model holding a commitment to them. Frequency is neither intention nor artistic worth.

The resulting fields nevertheless offer something to encounter. The first reader of S1-T2 interpreted its recurrent methodological concepts as a procedure that could make itself legible without yet arranging a consequential challenge. The reader explicitly questioned whether repeating “reciprocal legibility” could stand in for two genuinely distinct participants. Another reading remained possible: a cabinet of concepts offering invitations without prescribing their composition.

The reply enters; some of it cannot

Six fresh pretrained interpreter contexts—two per final field—made independent first readings and Carrier-field replies. Their first interpretations were saved before their replies. They had definitions available and retained their language training; they were not the small colour model and were not six authenticated, distinct model builds.

The replies then returned to the frozen small model as colour identities. No explanatory prose entered it, and its weights did not change. Each return was retained alongside a same-seed continuation from the original field. This makes numerical responsiveness inspectable.

The outcomes differed. R1, R2 and R4 changed all 24 sampled positions; R5 changed four. R3 and R6 produced the same 24 sampled colours as their counterfactual continuations, with only tiny first-step distribution differences. Those unchanged results remain in the gallery. This added diagnostic was specified after the predictive results were known and is an exploratory check of input sensitivity, not a measure of understanding.

It also exposes a boundary: the small model receives the reply’s colour sequence, but not its positions, typed links, membranes or deliberate blank cells. Those features remain visible to the next interpreter. They do not become understood merely because a new continuation appears. A reply can carry a distinction that this first model has no input channel for.

That is an unusually useful design lesson. A mental habitat needs a declared account of what its resident can notice. Otherwise the interface may appear to listen more deeply than the mechanism permits.

A habitat we can inspect

The public tool includes the actual selected weights and a local numerical runner. Choose starting colours, grow a field, pause, inspect the 64 hidden values, reveal the definitions when you wish, and save the result. No remote inference service receives the field. The state display visualizes numbers; it is not a window into subjective experience.

Chromatic Seed adds learned transition behaviour to Carrier’s expressive environment. It does not establish qualia, life, semantic understanding, or an advantage over other symbolic encodings. Its artistic interest remains open to the viewer, including the viewer who finds a field unconvincing.

There is now something smaller and more concrete than a promise: a model, its inherited environment, the patterns it acquired, the replies it received, and the parts of those replies it could not take in. That is a beginning we can work with.

Enter Chromatic Seed → Read the method and full evidence