J-space probes — Gemma 4 E4B

Interactive layer × position readouts of the model's global workspace (paper). Click any cell to pin a token and get rank-tracking charts. Pages are heavy (20–50 MB, fully self-contained).

→ Start with the interactive demo — guess the hidden emotion, explore the cross-model matrices, and play with the hallucination router. It's light and loads instantly.

• Multi-hop reasoning — "country shaped like a boot" → Italy in mid-layers, Euro at the top. The paper's flagship result, replicated.

The full layer × position deep-dive pages (covert fury/neutral, deletion threat, user grief, plus eight more probes) are 20–50 MB each and are being re-rendered in a lighter fetch mode before publishing. Everything needed to regenerate them locally: python probe.py --suite probes/emotions.json.