@memberjunction/rubrics

Server evaluators for rubrics. An evaluator produces answers. RubricScoring in @memberjunction/rubrics-base is the only 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.

Engine

sequenceDiagram
    autonumber
    participant Caller as Caller (action, test oracle, self-check, sampling job, CLI)
    participant Eng as RubricEngine
    participant Ev as Evaluator (registered by name)
    participant Srv as Evaluation entity server
    Caller->>Eng: EvaluateRecord(rubric, subject, context, evaluator or evaluatorConfig)
    Eng->>Eng: latest Published version (or versionId)
    Eng->>Eng: ResolveRubricEvaluatorSelection, then ClassFactory by name
    Eng->>Eng: content = input.content, or load the subject and ShapeContent
    Eng->>Ev: EvaluateRubric(version, content, settings, services)
    Ev-->>Eng: answers, and the prompt or agent run that produced them
    Eng->>Srv: save Draft evaluation, then submit the answers
    Srv->>Srv: RubricScoring.Compute — the only math
    Srv-->>Eng: score, outcome, per-criterion results
    Eng-->>Caller: { evaluationId, score, outcome, displayScore, criteria }
    Note over Eng,Srv: An evaluator error is stored as a Failed evaluation,
and EvaluateRecord throws with its message.
Copy

In server code, get an engine wired to a provider and a user:

import { ProviderRubricEngine } from '@memberjunction/rubrics';

const result = await ProviderRubricEngine(provider, contextUser).EvaluateRecord({
    rubricName: 'Research answer',
    subjectEntityName: 'MJ: AI Agent Runs',
    subjectRecordId: runId,
});
Copy

RubricEngine has its own Instance. It does not extend BaseSingleton; an engine is built per provider and user, so use ProviderRubricEngine.

Evaluate takes a version the caller already loaded and an evaluator name. It creates that evaluator with CreateRubricEvaluator, which looks the name up under BaseRubricEvaluator in the class factory, and calls EvaluateRubric with a context: the version, the subject content, the settings, and the services. It then saves a Draft, submits it, and returns the computed result. A throw from the evaluator produces a Failed evaluation with ErrorMessage. An unregistered name, or an evaluator a person completes (Human), throws before anything is saved.

EvaluateRecord also accepts evaluatorConfig, an evaluator selection in the AIAgentRubric.EvaluatorConfig shape. ResolveRubricEvaluatorSelection turns it into a name and settings; evaluator and settings on the call override them.

The services are what the engine lends an evaluator. ProviderRubricEngine supplies all three, bound to the caller's provider and user:

A call can replace any of them through EvaluateParams.services.

The draft stores the evaluator's EvaluatorType and EvaluatorName, the prompt or agent run that produced the evaluation in AIPromptRunID / AIAgentRunID, and Metadata.Evaluator: the name, the settings, and what the evaluator reported about its run.

RubricEvaluator.Evaluate(version, candidates) scores those candidates and returns { normalizedScore, rationale, evidence, result, answers }. BaseRubricEvaluator extends it and adds the plugin contract: EvaluatorName, EvaluatorType, IsAutomated, and EvaluateRubric(context).

Evaluators

Each registers under BaseRubricEvaluator by the name in bold. ListRubricEvaluators() reports every registered one. AI is an alias for Agent and AIPrompt for LLM.

• LLM (stored as AIPrompt) — three metadata prompts composed into one call: Rubric Evaluator (the parent, which owns the JSON reply contract), a judge in its judgePrompt slot (Rubric Evaluator - Default Judge unless PromptID or PromptName names another, for example Rubric Judge - Sage), and Rubric Criterion, which renders each criterion. SystemPromptID/Name and CriterionPromptID/Name swap the other two. The subject is a separate, nonce-delimited user message and never passes through a template. SinglePass is one call for the whole rubric; PerCriterion is one call per leaf. Samples runs it up to 9 times and keeps each criterion's median level. ModelID pins the model; ModelSelection: "Judge" lets the judge's bindings choose it. A quote that is not in the subject text is dropped.
• Decision (stored as AIPrompt) — every leaf with a level scale becomes a typed Score question, and all of them go to a Decision-type model in one call, Default Decision (Jev, then LLM Decision) unless a prompt is named. Each question is the criterion as Rubric Criterion renders it, after the rubric's instructions. The most probable level is the answer and its probability is the confidence. It refuses a rubric with an evidence-required leaf before calling anything, and leaves numeric-scale leaves unanswered.
• Agent (stored as Agent) — one run of the Rubric Evaluation Agent, or the agent AgentID names. That agent is a Loop agent. Its tools are Get Rubric and Get Rubric Subject. Get Rubric Consensus is not one of its tools. It does not publish.
• Deterministic — a rule on the criterion's evaluator config. No model call.
• Human — not run by the engine. HumanRubricEvaluator.Start creates a Draft evaluation and a task titled Score <rubric>. It does not score. StartHumanEvaluation constructs that class. The person answers in the form, and submit runs RubricScoring.

A custom evaluator subclasses BaseRubricEvaluator, registers with @RegisterClass(BaseRubricEvaluator, '<Name>'), and scores with this.Evaluate. The class factory constructs it with no arguments. Its own settings go in Settings.Extensions['<Name>']. The Rubrics Guide has a worked example.

Prompts are metadata

The engine writes no prompt text. promptData.ts builds the template data (Rubric, Mode, Criteria, Subject, and per criterion Key, Name, Levels with anchors, Hints, Scale, Text, …), and the prompts in /metadata/prompts render it with MJ's template engine. BuildCriteriaPromptData, PromptData, RenderCriteriaText, and BuildSubjectMessage are exported for custom evaluators that want the same contract. Sixteen judge prompts ship, the default and one per core agent, and every core agent's rubric link names its own. The Rubrics Guide covers the composition, the data contract, writing a judge, and previewing a composed prompt.

Register a provider for your own entity with RubricContentRegistry.Instance.Register(entityName, record => ({ text, data, files })); unregistered entities fall back to the record's readable columns. ShapeContent returns RubricSubjectContent: text, data, and files. A caller may pass content on Evaluate and skip the lookup. Test runs use input, expected output, actual output, and result details. Agent runs use the final payload and the in-memory message. There is no turns column and no transcript column.

Actions

Each action calls RubricEngine. None of them call another action.

Publishing a version is the publish path, or the publication metadata for the eight shipped agent rubrics. Create Rubric Draft, the architect import, and mj rubric evaluate do not publish.

CLI

mj rubric is a shim in @memberjunction/cli. The commands are implemented here, in @memberjunction/rubrics. The shim opens the provider and closes it.

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>] [--prompt <name>] [--model <id>] [--mode SinglePass|PerCriterion]
Copy

--evaluator takes any registered evaluator name and defaults to LLM. --prompt names the judge for LLM (for example "Rubric Judge - Sage") or the decision prompt for Decision; --prompt, --model, and --mode become the evaluator's settings. evaluate does not publish.

mj test run --rubric <name-or-id>[@version] and mj test suite --rubric <name-or-id>[@version] pin that rubric for the run. The version is 1.2.0 or a version id. Resolution order is the run flag, the rubric oracle's own config, Test.RubricID, the suite's RubricID walking up parents, then the agent's default Evaluation rubric.

mj test promote-criteria <test> copies the test's inline judge criteria onto a Draft rubric and sets Test.RubricID. The version is not published. A test that already has an llm-judge oracle keeps that judge when the rubric choice came from the agent.