SnailSafe is a small research lab studying what happens before it does.
The final text is the visible end of an internal process. This page is about the work of making that process observable — and about the people and models doing it.
It began with a question: what happens inside an AI model before it commits to an answer? Not a product category. Not a market segment. A question that repeated itself across different models and different tasks.
The first observations were incidental — trajectories that took shape before the text appeared, timing differences between models on the same probe, coupling signals that arrived before commitment. As they accumulated, they suggested that inference might be studied not only as internal structure, but as a process unfolding over time.
SnailSafe’s findings are empirical and bounded by the models, tasks, instruments, and diagnostic conditions under which they were recorded.
Different model families do not always expose inference dynamics through the same measurement channel. Tokenizer geometry, architecture, reasoning style, and internal representation can affect what becomes readable and when.
Background spans traveling-wave tube engineering, large-scale computing systems, product development, and independent AI research. SnailSafe emerged as an outgrowth of that work — not from a plan, but from a question that kept returning.
The rest of the “team” is a cohort of AI models used as instruments, drafting partners, and cross-checkers. They run diagnostics, review findings, help compose methods documentation, and — transparently — assisted with this website.
This is a small operation by design. Findings are validated slowly. Claims are scoped to what has actually been measured.
Every observation begins with a question.
We respond directly, usually within a day or two.