license: agpl-3.0
language:
- en
base_model: Qwen/Qwen3-8B
tags:
- interpretability
- attention-schema
- experimental
Strange loop in J-space: the c8 reader
Wiktor Tomasik 路 documentation updated 3 November 2026
This download contains one compact reader checkpoint, not Qwen3-8B itself. The reader tensors remain unchanged and exactly match the reader in the new recorded experiments. The base language model and the reader's neural weights stay frozen during development; stance, habit and attention-prediction statistics develop online.
The reader selects a sixteen-observation window of layer-24 Qwen activity. Its eight-number stance, learned habit and attention forecast affect selection. A 32-number picture supplies the direction used to turn stance. The forecast predicts selected-minus-average key content; it is not an explicit record of a proposition or its ownership. Live operation also writes a wide native content contrast before layer 16 and reads the processed return after layer 24.
The current evidence separates a private way of selecting from a mostly shared compact picture. Interpretation-to-attention feedback is a later experimental route: it changes recorded organization with a cost to text responsiveness. These variants are not new default weights, and their language behavior is not established by the earlier reference tests. A complete fine-tuned J-workspace/self-model fusion and subjective awareness are not demonstrated.
Read the unified study 路 Scientific report and code 路 Numerical evidence 路 Companion Space.
Run the authored recorded demo
Install PyTorch for your hardware and NumPy, then run:
python reproduce.py --states demo/states.npz --out results/demo --steps 468 --warmup 0 --device cpu
The four identical histories should preserve symmetry; this is a numerical check, not an individuality demonstration. A developed snapshot includes controller state and observations but does not include a Qwen context/cache. Original external passages are omitted. Software and reader weights retain AGPL-3.0-only; new author-owned research text/figures/data retain CC BY-SA 4.0, with upstream terms preserved. The c8 reference and current documentation are publicly available.
Reuse and citation policy: discovery and human-led contributions are welcome. The author requests no general-purpose model training and official upstream contributions only for automated research development. Existing open licenses remain valid.