CHROMATIC SEED / JOB 134 / PROTOCOL 1
Frozen before training and before test evaluation. This is a local, author-designed demonstration with one trained model, not a preregistered external study.

QUESTION
Can a randomly initialized small recurrent model learn predictive structure from numeric Carrier colour identities and their existing graph, then generate inspectable fields that separate interpreters can encounter and answer?

INPUT BOUNDARY
Use the frozen Job133 checked palette (6,257 colours) and its directed topology. Concepts' titles, definitions, prose, past artworks, human interpretations and pretrained weights never enter this model's training arrays. Colours are discrete identities: their RGB distances have no assumed semantic meaning. The supplied graph embodies prior human/AI editorial structure; stripping words does not remove that inheritance. A 32-dimensional embedding, 64-state one-layer GRU and categorical output head are initialized from random numbers. This is next-colour prediction, not a demonstration of subjective sensation or a general thinking agent.

SPLIT AND DATA
Group each directed edge with its reverse under an unordered pair; SHA256 of '134-edge-v1:minIndex:maxIndex' maps pairs to 80/10/10 train/development/test intervals. Integer indices follow checked ID order. Random walks use only training edges, 32,768 sequences of up to 16 targets. Ended walks use ignored targets, not new tokens. Held-out queries predict the endpoint of a development/test edge after up to eight preceding colours connected solely by training edges. All held-out directed edges are scored, including rare endpoints; no eligibility filtering after seeing results. This tests one graph-edge completion task on one vocabulary, not semantic truth, artistic quality or generalization to new concepts. All graph pairs and deterministic seeds are archived.

FIT AND MODEL SELECTION
One initialization seed, one data seed, one run. Fixed 1,200 updates, batch64, AdamW lr0.002, weight_decay0.0001, gradient clipping1.0; evaluate initial and every200 updates on development negative log likelihood. Select minimum development loss, with earliest step for a tie. No final-test input is read by training. Save initial, selected and final parameters, development history, versions, wall time and resource records. No test-driven tuning; no extra training if the chosen model performs poorly.

HELD-OUT CHECK
Once the chosen checkpoint is fixed, compare all test targets among the full vocabulary with uniform probability, training-target frequency with add-one smoothing, a smoothed outgoing training-edge distribution, a smoothed two-hop training-graph distribution, and the same untrained model. For graph baselines use 0.9 transition probability plus 0.1 uniform; missing outgoing mass falls back to uniform. Two hops means two uniform neighbour draws in the training graph only. Report mean negative log probability, top1/top10 accuracy and reciprocal rank, averaging each metric over a uniformly random ordering within tied groups. These are descriptive scores from dependent graph queries, not independent participant trials; do not use a query bootstrap as a population significance test. Publish all baseline and model scores, including failure. No claim about semantic understanding follows from edge prediction.

GENERATIONS
Three random four-colour prefixes and all sampling seeds are fixed in generation.json before fitting. Temperature0.9, full categorical sampling, no top-k, rejection or repetition penalty. Each stream supplies three successive 24-colour fields, using up to64 immediately preceding indices as context. Generate matching initial-model controls for all three streams, retaining the first field of each. Positions follow a fixed six-by-four serpentine grid; z=0 and emphasis=2. Links and regions are empty. The model chooses colour identities; the adapter supplies geometry, addresses, state boundaries and grammar. Save every field, sampled probabilities, full logits and hidden-state trajectories. There is no concealed English source. Duplicate selections remain visible.

ENCOUNTER AND RETURN
Two fresh pretrained interpreter contexts independently encounter the final learned field of each stream (six encounters). They receive its image, raw field/codec, selected checked definitions and grammar, but no other interpretation, no prior authored artwork, no held-out scores and no training graph. They are told the small model selected colours and geometry is fixed; they are not blinded about the medium. They write a first interpretation with explicit alternatives and anchored cells, save it before constructing one Carrier-field reply or a reasoned refusal, and disclose their own pretrained language knowledge. Agent file-scope limits are instructions, not an operating-system sandbox. These are six outputs from the available agent system, not six authenticated independent model builds and not a human reception study.
The reply's non-null colours in stored cell-array order become the frozen small model's next context (up to64); prose, labels, geometry and new relation types do not enter it. A separate fixed seed generates24 return colours. No online training occurs and no semantic comprehension of the reply is inferred. Preserve original interpretation, reply and return, including disagreement/refusal/failure. Root may provide its own later interpretation, labelled as such.

PUBLIC TOOL
Expose the selected small model locally in the browser with manual seed colours,24-step generation, downloadable raw fields, optional definitions and a64-number hidden-state view. A hidden-state picture is a visualization of numeric activations, not an image of experience. Compare browser-JavaScript logits with the Python model numerically; floating point and sampling near cumulative boundaries need not be bit-identical. Report actual UI checks and any untested browser/phone/live-site behavior.

INTERPRETIVE LIMITS
Training loss/edge scores, exact transport and artistic reception are different evidence. An interesting reading cannot validate model understanding; a disappointing edge score cannot decide whether an artwork matters. The learned model uses only the palette and supplied topology in its training data, but creator, curator and interpreter decisions surround it. The colour allocation is an arbitrary shared address system, not calibrated qualia. The only added learned behaviour claimed here must be supported by weights and measured output changes.
