@memberjunction/rubrics-base

UI-safe scoring and version classification for rubrics. Pure functions: no database and no Angular. The server and the author preview both call this package. Neither one keeps a second copy of the math.

Start with the Rubrics Guide — the model, the five common tasks, tests, agents, and adopting rubrics in an application. This README is the package reference.

Scoring

RubricScoring.Compute scores a version snapshot and its answers. A levels answer uses the level's normalized value. A numeric answer maps the raw value across the scale's minimum and maximum, and inverts that when lower is better. A Percentage scale is a numeric scale named Percentage, or any numeric scale whose minimum is 0 and maximum is 100.

The outcome is the first match:

1. Incomplete when completeness is below the version minimum.
2. NotApplicableFailure when a not-applicable answer used Fail the evaluation.
3. Incomplete when the overall score is null.
4. GateFailed when a non-advisory gate is unanswered or below its gate minimum. The minimum is 0 when the criterion does not set one.
5. Passed or BelowThreshold when a pass threshold is set.
6. Scored otherwise.

Passed is true only for Passed. It is null for Scored, and for Incomplete when the version has no threshold and no gate. Comparisons use the score rounded to six decimal places.

Versions

RubricVersionDiff.Diff compares a draft with the version it was cloned from. The bump says whether scores stay comparable.

• Major — a non-advisory node was added or removed, or a scoring field on a non-advisory node changed, or IsAdvisory flipped, or the version's not-applicable policy changed, or a scale used by a non-advisory node changed its type, range, step, direction, or a level's value. A node that is advisory on both sides does not make those scoring-field edits Major.
• Minor — the pass threshold or minimum completeness changed, a band was added or removed or its range or tone changed, or an advisory node's scoring fields changed, or evidence or rationale became required.
• Patch — wording, sequence, and the display range. Bands match by label. A renamed label on the same band is Patch.

HighestNonDraftVersion is the highest Published or Retired version. A Draft is not a base. The first publish, with no such base, is 1.0.0. The author may request a higher bump and may not request a lower one. A draft identical to its base cannot be published.

SnapshotFromRows builds the snapshot those functions score. EvidenceJson turns plain text into a quote list and leaves an evidence array as JSON.

BandFor names the band for a 0..1 score. WeightShares is each node's percent of its non-advisory siblings. DraftProblems names a tree that cannot be published. Frozen is true when a published version uses the scale. The author widgets call these. They do not keep a second copy.

CLI

This package does not talk to the database. The commands that do are:

mj rubric list
mj rubric show <rubric>[@version]
mj rubric diff <rubric> <version> <version>
mj rubric validate <file>
mj rubric evaluate --rubric <rubric> --entity <name> --record <id> [--evaluator <name>]
mj test run --rubric <rubric>[@version]
mj test suite --rubric <rubric>[@version]
mj test promote-criteria <test>
Copy

mj rubric evaluate defaults to LLM, which is stored as AIPrompt. None of these commands publish a version. mj test promote-criteria copies a test's inline judge criteria onto a Draft rubric and sets Test.RubricID. @memberjunction/rubrics is the server package that runs an evaluation.