Manifesto
The right to your own intelligence
People should be able to inspect the AI systems they rely on, including the models, records, and work those systems perform.
Owning a model file is useful, but possession alone is not control. A person also needs to know where a system runs, what it remembers, who can change it, and how an answer was produced. Many AI products remove those facts from view. Prompts cross a network, models change behind an API, and work disappears into terminals that do not share a record.
This problem is not limited to hosted models. A local system can still be opaque, and a team of coding agents can still lose decisions, overwrite files, or stall without telling the operator. The relevant question is whether the person responsible for the result can inspect and intervene in the system that produced it.
Torad's compact-model research asks how much capability can fit on hardware one person can own and examine. Smaller models are not automatically better, but the constraint makes memory, training method, representation, and runtime architecture visible.
Fleet applies the same standard to multi-agent work. Claude Code, Codex, Grok Build, and Gemini CLI remain in their native tools, while Fleet records ownership, decisions, liveness, and recovery in one place.
Inspection does not require every model to run locally or every component to be open source. It requires clear boundaries. A measured result must show its comparison. An unpublished method must be labeled unpublished. A coordination boundary must not be described as a security sandbox. A beta must not be presented as a finished release.
Torad will use small local models where they are sufficient, strong remote models where they are useful, and coordination software that keeps a person responsible for the outcome. The system should expose enough of its operation to be questioned.