CHROMATIC SEED / JOB134
Open publication/index.html for the local instrument; publication/posts/index.html is the WordPress copy desk.
Upload publication/upload contents to the website document root. Destination /carrier/chromatic-seed/.
The browser model is local; weights and definition data are page assets. No API or paid service is needed.

PUBLICATION FILES
Four Custom HTML blocks: tool, news, blog, reflection. Standalone previews and unstyled bodies are also included.
No site deployment was performed. Physical browser/phone/theme behavior is not claimed by the inert checks.

REPRODUCTION
Preserve this record. Work in a separate copy. Portable inference needs Python and NumPy only:
python tools/infer_numpy.py --prefix 0 25 36 37 --seed 134100
The example prefix contains model indices, not an English sentence.
For training, use Python3.12, NumPy2.3.5 and PyTorch2.7.1. The retained run used CUDA12.6 on RTX3070Ti Laptop.
The recorded common/runtime files provide the pinned dictionary/lexicon/codec; no live registry access is needed.
In a separate copy, relocate runs/seed01 to preserve it before invoking tools/train_seed.py, which refuses overwrite.
Run tools/evaluate_seed.py after training, in that separate copy with its evaluation outputs relocated as well.
Evaluation data is now public; it must not become a supposedly hidden benchmark for future tuning.
tools/generate_fields.py uses retained numeric prefixes; fresh output directories prevent replacing original fields.
The receiver codec requires Node. Example:
node common/field.cjs decode artworks/S1-T2/S1-T2.png --layout artworks/S1-T2/S1-T2.layout.json --out a-new-recovery
Optional model tests: tools/test_seed_model.py. Figures: tools/figures.py with Matplotlib3.10.7.
Generation uses xorshift32 and full softmax at0.9. Floating point may differ across hardware; sample equality is not promised universally.

INPUT/OUTPUT BOUNDARIES
The GRU consumes categorical colour identities, not RGB channel geometry or natural-language definitions.
The small model cannot take in reply links, regions, nulls, position, depth or emphasis. Those remain in the public field records for interpreters.
Interpreters are separately pretrained language agents. Their frozen readings and replies do not certify the small model's understanding.
Raw images contain exact concept-colour cores and an embedded CVP2 packet with different byte-transport colours.
Source dictionary attribution/provenance remains in common/runtime/dictionary.json. No new blanket licence is asserted over inherited materials.
