BlackWindow: watch a frozen model be read, live.

The Sunstone North Lab is a live instrument, not a slideshow. Every answer on this page is generated on demand by a frozen 31-billion-parameter multimodal language model served from one desktop computer, read through small trained artifacts from the SRT (Semiotic-Reflexive Transformer) research program.

01 · Generate

Chat with gemma-4-31B in 4-bit. Tokens stream live, tinted by the model's per-token uncertainty, charted as they land. The hardest steps are read back as captions through a 44 MB linear head. Turns share a session with KV-cache memory, so history never re-prefills.

02 · Read an image

A picture retrieves its own words: the native VLM caption plus nearest captions through the head, from one forward pass. Uploaded images are processed in memory and never stored.

03 · Find images

Words retrieve pictures. Your text is placed in head space and matched against the full COCO corpus, more than 122,000 images, with one matrix multiply.

04 · Stand somewhere

A map of what the 31B has read. 40,000 COCO photographs, sampled from a 123,287 image gallery, are laid out by how gemma-4-31B read them, and every region name on the map was written by a frozen 382 MB Qwen3-0.6B from that region's own centre rather than chosen by us. Click anywhere to read what is there, type a sentence to see where it lands, or drop in a photograph: the 31B encodes it here, on this machine, into the same 1,024 dimensional space the gallery was built in, so it lands among its real neighbours. Measured on 500 held-out photographs against 118,287, the reader's sentence retrieves the picture it was written from at median rank 64, against 45 for a human caption the head never trained on, and the same reader handed another photograph's point scores at chance.

Consensus

A coding model answers the same request eight times. The selector runs all eight on inputs invented from their own signatures, keeps the largest group that agrees, and returns one, with no tests, no reference solution and no training. Ask for a Python function live: generation and execution happen on a disposable GPU box, each candidate as an unprivileged user with no writable disk, and nothing is stored. Beneath it, banked pools from six model sizes on 164 HumanEval problems carry the ground truth, so the selector's misses are as visible as its hits.

Introspect and Steer

A second workspace runs live introspection on a frozen Qwen2.5-7B: per-token divergence, reflexivity, regime, and hidden states read back as text. The steering bench injects a meaning direction built from two sentences into the residual stream with a dial, with a baseline lane for honest comparison.

Built by Sunstone North, the engineering and product studio of the SRT research program. Code and checkpoints: GitHub · Hugging Face.

This page has a full interactive version; enable JavaScript to view it.