A linear probe on the hidden state predicts whether an answer is right better than the model's stated confidence does, in 12 of 16 cases across four open models. On one model, erasing that direction collapses the stated confidence too. The model knows more than it says, and sometimes uses it.
Can a language model draw a geometry diagram? The model describes the construction, an engine draws it, and every prompt ships with its own key, so a compiler grades the result instead of a vision model. Graded this way, the scaffold matters more than the model.
Consciousness prompts against matched philosophy controls in a small reasoning model. A late-layer direction separates them almost perfectly, in a direction so low in variance that PCA never shows it. A scatter plot of this data would have produced a false null.
Introspective grounding. When a model says it is uncomfortable, is anything inside it actually changing? I nudge a candidate direction up and down, keep everything else fixed, and watch whether the words follow.
Welfare signal dissociation. Do what a model does, what it says, and what shows up inside it line up? I am mapping where they agree and where they quietly split, across models.
Simulation observatory. Small social games between AI agents with hidden incentives. I keep each agent's private notes, predictions, and public words side by side, and watch what happens when they drift apart.
The first three use the real numbers and activations behind the papers. Nothing in them is simulated.
Brain-computer interfaces, 2017 to 2019: tactile EEG across age groups, paper; mindfulness and BCI accuracy, paper; music imagery as a control signal, paper.
