Evidence-first assessment · documentation

Assess a small evidence bundle and inspect its receipt before acting on a score.

Hermes Rubric asks a configured model backend to synthesize criteria, collect cited evidence, and score against that evidence. Start with a disposable two-file bundle so each output field has a clear source.

View source ↗Open-source catalog

§ 01

Install and verify

Pinned first workflow

Start by checking the released command is available.

$ uvx --from hermes-rubric==1.2.3 hermes-rubric --version

§ 02

Run one bounded task

Read the result

$ bundle=$(mktemp -d)
printf '%s\n' 'The staging migration updated every order row.' > "$bundle/review.md"
printf '%s\n' 'Staging log: 12,004 of 12,004 rows updated; no production run.' > "$bundle/evidence.md"
uvx --from hermes-rubric==1.2.3 hermes-rubric --intent 'Assess whether the review is supported by this bundle.' --context "$bundle/review.md" --target "$bundle" --target-type review-bundle --out "$bundle/assessment.json" --verbose
cat "$bundle/assessment.json"
rm -rf "$bundle"

Expected result. With Claude Code or a local Ollama backend available, stderr shows rubric synthesis, evidence collection, and scoring, then assessment.json contains citations, coverage, hedge fields, and a receipt. Without a backend, it exits 1 and writes no result.

§ 03

Reliance boundary

The aggregate is a signal, not a release decision. Criteria are model output unless you supply a frozen rubric; partial coverage means material may be uninspected, not absent.

Continue with the canonical repository reference; this route follows the published source and does not replace it.