Skip to content

7. Submitting, and where next

Submitting

\usepackage[final]{vouch} removes the links and the provenance appendix, and leaves plain numbers. vouch check --strict still verifies everything, so keep it in CI.

Where next

  • Numbers you don't have yet. vouch.expect("key", desc=..., producer="python ...") in vouch_values.py lets you cite a result before the experiment exists. The PDF shows [pending: key] and vouch todo lists what's owed. See Derived values, claims, tables.
  • Comparisons. vouch compare evaluate.knn.n_train_640.acc evaluate.linear.n_train_640.acc prints the difference, a Welch t-test and the @vouch.derive and @vouch.claim that make it citable. --write adds them to vouch_values.py. See vouch compare.
  • No code changes at all. List the function in vouch.toml ([[track]] function = "experiment.py::evaluate") and run python -m vouch.exec experiment.py. See Recording (Python API).
  • Writing with an LLM agent. vouch init --agents installs a Claude Code skill and a hook. The hook catches a typed number or a mistyped key the moment the agent writes it. See Claude Code.
  • The full design is in SPEC.md in the repository.

You've now covered the whole loop: record, find, cite, derive, check, and respond to a change. From here, the Guide and CLI reference sections go deeper on each part.