2. Describe the metrics, then find the key

Describe each family once

vouch asked for descriptions. Describe each family of keys once, in vouch.toml:

[metrics]
"*.acc"       = { fmt = ".1pct", better = "higher", desc = "test accuracy of {1} ({2}), mean and std over seeds" }
"*.train_acc" = { fmt = ".1pct", better = "higher", desc = "training accuracy of {1} ({2}), mean and std over seeds" }

fmt = ".1pct" prints 0.873 as 87.3%, and {1}, {2} are key segments. You don't need to re-run anything: formats and descriptions are applied when values are read, not when they are recorded.

Find the key you need

$ vouch explore --open

This opens a local page listing every value by script, then function. One click copies the LaTeX that cites it.

vouch explore --open: the left side lists scripts and functions; the right side lists that function's keys, each with a button that copies its citation.

From the terminal:

$ vouch search "knn test accuracy 640"
evaluate.knn.n_train_640.acc           87.3 ± 1.4%   test accuracy of knn (n_train_640), mean and std over seeds   (experiment)
evaluate.knn.n_train_640.train_acc     87.4 ± 0.9%   training accuracy of knn (n_train_640), mean and std over se   (experiment)
...
$ vouch cite evaluate.knn.n_train_640.acc
\vouch{evaluate.knn.n_train_640.acc}         →  87.3 ± 1.4%    (fmt .1pct)
\vouch[.2pct]{evaluate.knn.n_train_640.acc}  →  87.30 ± 1.41%    (fmt .2pct)
test accuracy of knn (n_train_640), mean and std over seeds · higher is better · fresh · run experiment
subfields: .mean 87.3% · .std 1.4% · .n 5 · .ci95 85.5-89.1% · .min 85.8% · .max 89.1%

After the first vouch build, .vouch/CATALOG.md lists every key on one line each. It's the file to give an LLM agent that is writing the paper with you — see LLM & agents.

Full reference: vouch explore, vouch search/vouch cite.

Next: 3. Cite and build.