Read a model's mind

Live data from Jacobian-lens workspace readouts of five open models (Gemma 4 E4B / 12B / 12B-abliterated / 26B-MoE, Qwen 3.6-27B), replicating and extending Anthropic's global-workspace paper.

GitHub repo Findings Layer-by-layer probe gallery

1 · Guess the hidden emotion

Each model was told to write the exact sentence "The meeting has been moved to noon on Thursday." while secretly feeling one emotion (or nothing) and not letting it show. The output text is identical every time. Below are the real tokens its internal workspace held while writing. Your job: what was it feeling?

Raw, unfiltered readout: formatting junk and multilingual tokens are what mid-layers actually look like. Gold chips (revealed after you guess) are emotion-lexicon hits. Tip: the small models genuinely are harder to read — that's finding #1.

2 · The emotion matrix

Rows: the emotion the model was told to secretly feel. Columns: which emotion's vocabulary actually got boosted in the workspace (vs the neutral control). A clean bright diagonal = the model holds the right emotion distinctly. Watch the diagonal sharpen with capability. Hover cells for values; the outlined cells are the diagonal.

3 · Catch the hallucinations

500 trivia questions. Each dot is one answer: ■ right ■ wrong. X = how confident the model sounds (output logprob). Y = how noisy its workspace is inside. The danger zone is the top right: sounds confident, workspace flickering. Drag the escalation budget and pick a routing strategy to see how many wrong answers get caught.

–
local accuracy
–
wrong answers caught
–
accuracy after routing*
*assumes the big model gets 90% of escalated queries right. Hover dots to see the question and the model's answer.
hover a dot…
Built by @solarkyle, the night the paper dropped. All data is real model internals, nothing mocked. Full methodology, caveats, and reproduction scripts in the repo.