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 ...")invouch_values.pylets you cite a result before the experiment exists. The PDF shows[pending: key]andvouch todolists what's owed. See Derived values, claims, tables. - Comparisons.
vouch compare evaluate.knn.n_train_640.acc evaluate.linear.n_train_640.accprints the difference, a Welch t-test and the@vouch.deriveand@vouch.claimthat make it citable.--writeadds them tovouch_values.py. Seevouch compare. - No code changes at all. List the function in
vouch.toml([[track]] function = "experiment.py::evaluate") and runpython -m vouch.exec experiment.py. See Recording (Python API). - Writing with an LLM agent.
vouch init --agentsinstalls 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.