Evaluation
Try a contract against code that follows it, then make a small change that breaks it.
tenet check --snapshot ./project
tenet check --snapshot ./project --patch ./break-cache.patch./project holds the code and its .contracts, and the patch makes one small change that breaks a contract.
The first command checks the whole project. The second assumes the original project satisfied the
contracts and checks whether the patch breaks them. Each run uses an isolated copy.
Keep .contracts inside the project. Use ordinary code examples in the contract to explain the rule;
they are documentation, not a separate test language.
To evaluate an agent workflow, have the agent review the project with its ordinary tools and Tenet. Grade the final review against the code and a separate rubric. Keep grading instructions outside the agent's checkout.
Compare a plain review with one that receives contracts and Tenet. Use the same model and budget, and give each run fresh history. This measures the combined effect of contracts and tools; it does not isolate the CLI's contribution. Include allowed changes, violations, missing context, and requirements spanning several files.
Score missed defects, invented findings, useful explanations, and appropriate uncertainty. Track project-specific rules separately from general correctness. Save tool traces to understand why the agent missed something. A model grade helps compare runs; it does not prove correctness. Keep the judge and rubric fixed when comparing changes.