Author
Make the proposal.
External models can write the code Garnet inspects. Acceptance rests on the tool output and the reviewer’s decision, rather than the model’s explanation of its own work.
Research / Decision models
Models can write.
Evidence must hold.
Where should a model sit in a language built for the code agents write and humans accept?
A place for intelligence.
A boundary for authority.
- 01AuthorA model proposes code.
- 02InspectCheck source and declared authority.
- 03DecideA person considers the evidence.
- 04RecomputeVerify the recorded source binding.
Conceptual workflow. A content seal binds source; it does not establish authorship or prove a human decision.
Three roles. Different responsibilities.
Adviser
Raise the question.
The research proposes models for finding triage, review priority, and spotting overclaims. An advisory result can raise an item for review; it cannot waive a deterministic failure.
Governed effect
Explore the next boundary.
A model call inside a Garnet program would need explicit authority, isolation, and replay rules. This is a research direction, separate from the current language surface.
Questions worth testing
Useful advice.
Recomputable verdicts.
The compilation considered Laya and Jev as decision-model candidates. It asks whether constrained outputs can help people order reviews and identify surprises without entering Garnet’s deterministic trust path.
Candidate names describe the research considered. They do not announce integrations, measured Garnet results, or approved product features.
Research method and scope
The internal source combines video analysis, sampled frames, repo grounding, competing options, and skeptical review. This public digest presents its research questions and recommendations. No model was run for that compilation, and this page reports no reproduced benchmark or new runtime capability.