Watch a frozen model be read.

The Sunstone North Lab walks you through the SRT instruments chapter by chapter, using real precomputed artifacts from frozen language models. Nothing is generated live, and nothing needs to be: reading was done once, and replaying it costs nothing.

The chapters

• Reading a mind mid-sentence: a frozen Qwen2.5-7B answers while a 12M-parameter instrument reads divergence, reflexivity, regime, and ten hidden states back as text.
• A picture retrieves its own words: a text-trained read-out on a frozen multimodal model names what it sees, zero image training.
• Steering the stream: a meaning direction built from two sentences, swept over eleven real generations, with a factual control that never breaks.
• Same head, different silicon: the substrate-invariance result, from a 10× host reduction to 4-bit weights to Apple Silicon at 100% agreement.

The links above are complete all-HTML versions of each chapter. The interactive walkthroughs require JavaScript. Corporate site: sunstonenorth.com · Research: github.com/space-bacon/SRT · Weights: huggingface.co/RiverRider