The work began with an observation.
An AI model’s answer can have a measurable internal history before it becomes text. These papers document the first stages of the SnailSafe research program — from evidence that governability differs sharply across model architectures to a physical framework for measuring inference dynamics and pre-commitment signals.
Two papers document the path from silent failure to measurable inference dynamics.
Each is an arXiv preprint by Gregory M. Ruddell. Together they establish the motivating problem, define governability, and begin to characterize the regimes that determine whether a warning signal appears before commitment.
The first paper establishes the problem.
The second begins to describe its geometry.
Together they mark the transition from observing silent failure to building an instrument capable of characterizing when inference becomes measurable.
The papers are evidence of a research program — not a claim that every question is settled.
These findings are empirical and bounded by the models, tasks, prompts, instruments, and inference conditions under which they were recorded.
The papers show that pre-commitment observability and governability can be measured in specific regimes. They also show that the same measurement does not read every model cleanly.
That boundary is not separate from the research. It is part of the research.
Every new observation begins with a question.
SnailSafe welcomes conversations with model developers, research laboratories, evaluation teams, and organizations investigating inference-time behavior.