TARS AI
Local model and Mac runtime research
Local model experiments connected to a private Mac runtime, persistent workrooms and explicit evaluation gates.
- Status
- Research prototype
- Kind
- AI & systems
- Stack
- Local models · SQLite · Evaluation tooling
BUILD NOTES / Research prototypeTARS AI
Project overview · not an interface capture
Local model and Mac runtime research
Local models / SQLite / Evaluation toolingProject overview · not an interface capture
The problem
A local assistant needs evidence of retained behavior and safe tool use before it can be trusted to act.
What was built
- A private Mac runtime with SQLite-backed persistence and bounded workroom tools.
- Training experiments, tool-decoder evaluation and retention checks.
- Regression-based candidate selection that rejects failing model candidates.
What I learned
An improved individual benchmark does not justify accepting a candidate that regresses established behavior.