CHROMATIC SEED / PORTABLE NUMERIC MODEL

This smaller ZIP contains the selected weights, numeric palette and two inference implementations. It does not contain the full training data, artworks, website or receiver codec. Those are in full-record.zip.

Python3.12 and NumPy2.3.5 were used. PyTorch is not required for inference:
python tools/infer_numpy.py --prefix 0 25 36 37 --seed 134100

The prefix contains categorical model indices into data/palette.json. These indices need not be interpreted as current online Carrier IDs. The output is24 sampled colours and the final64-number hidden state.

runs/seed01/best.npz holds seven plain numeric arrays. tools/seed_model.py implements a NumPy path and, if PyTorch is installed, the training model. It does not download weights.

publication/model.js holds the same arrays as little-endian Float32 base64 for publication/inference.js. The JavaScript runner has a CommonJS export and a browser global. Its recurrence was compared numerically with the trained PyTorch model. Individual samples can differ near floating-point sampling boundaries on other systems.

This is a frozen small sequence model trained from random weights. It does not read definitions, positions, typed links, regions, blank cells or an English interpretation. Numeric state is not evidence of subjective experience.

For the complete data, artwork, method, attribution and frozen evidence see the full record or https://www.unifiedstate.us/carrier/chromatic-seed/method.html .
