Jsort

It reports itself as one. With -o each line carries a score and a standard error, and a run prints reliability and a first-position lean so you can see order bias: 51 FOMC sentences gave 255 comparisons in 10 rounds, reliability 0.96, first-position lean -0.07, $0.0037 in 7.2 seconds.

Comparisons, not characters. A thousand short lines take about 5,000 comparisons, roughly a minute and seven cents. It asks about a few pairs per text rather than all of them (-k 10 by default, minimum 2) and then fits the scale, so cost scales with the number of texts and the rounds needed.

Write it as a comparative, since it goes into one sentence: Text A ranks higher than text B on this criterion. The README does not publish an accuracy figure for the fit in the digest available here, so treat the reported standard error as the check: a large se on your own data means the description is not discriminating.

It records its provenance and refuses to mix. --save-scale writes the fitted scale for --scale to place later texts on, describing each new text against the scale's anchors, and every answer is stored with the model that gave it. Placing against a different API, endpoint or model needs the explicit --any-model flag.

Run uv tool install jev-sort. It finds a key the way jgrep does, from TYPESAFE_API_KEY, OPENROUTER_API_KEY or a System One gateway, in the environment or under ~/.config/jev/.

About 5,000 comparisons, roughly a minute and seven cents. All JevKit tools share one cache file, so repeated runs with the same seed come back from cache.

Yes. jsort --save-scale writes the scale to a file, and jsort --scale places later texts against its anchors, though --any-model is needed if the API, endpoint or model differs.