Skip to content
FlowMCP v4

Scoring Protocol

Specification for grading FlowMCP v4 schemas via LLM evaluation. Documents the data formats exchanged between the CLI and an external Grader (e.g. Claude Code harness, third-party implementation).

Purpose

Section titled “Purpose”

The scoring protocol decouples three concerns:

  1. Prompt generation — done deterministically by flowmcp-core/v4/GradeReporter.buildEvalPrompts
  2. Score production — done by an external Grader (LLM-based)
  3. Grade calculation — done deterministically by flowmcp-core/v4/GradeReporter.grade

CLI is responsible for 1 and 3. Grader is responsible for 2. Communication via JSON files (prompts.json, scores.json).

Delegation — the grading model lives in the Grading-Spec

Section titled “Delegation — the grading model lives in the Grading-Spec”

FlowMCP is the highest instance and delegates the grading model to a separate, independently versioned standard: the Grading-Spec (gradingSpec/3.0.0), published as a dedicated docs area. The Grading-Spec owns the grading model — score-to-grade thresholds, the extended dimension set, Scoring System, Grading System, Categorical-Veto, Tier-trim, and skill families.

This file owns the upstream scoring transport only: the deterministic prompts.json / scores.json artefact pair exchanged between the CLI and an external Grader. The Grading-Spec sub-consumes this transport and treats it as the highest instance for the artefact pair — it does not redefine it.

Entry point: Grading-Spec 3.0.0 — Overview.

Protocol Version

Section titled “Protocol Version”

scoringProtocol: "v1"

Versioning allows future changes without breaking existing graders. Implementations MUST check the version and fail gracefully on unknown values.

Dimensions

Section titled “Dimensions”

Each schema is evaluated on 2 dimensions:

DimensionWhat is rated
whenToUseClarity and specificity of the schema description (does an LLM know when to use this schema?)
parametersHow well the parameter descriptions enable correct tool invocation

Each dimension yields a score on a 1.0-5.0 scale (floating point). These two are the base transport dimensions; the Grading-Spec (08-grading-model.md §Dimension Enum) extends this set with the additional grading dimensions it owns.

Grade Thresholds — delegated to the Grading-Spec

Section titled “Grade Thresholds — delegated to the Grading-Spec”

The score-to-grade banding, the production gate, the tier-trim rule and the Categorical-Veto list are owned by the Grading-Spec (gradingSystem/1.0.0), not by this transport protocol. See Grading-Spec 3.0.0 — §4.1 Score-to-Grade Thresholds. Changing any threshold or the numeric mapping bumps the gradingSystem version independently of scoringProtocol.

File Formats

Section titled “File Formats”

prompts.json — Emitted by CLI

Section titled “prompts.json — Emitted by CLI”

Written by flowmcp grading non-deterministic <schema> --emit-prompts. Read by Grader.

{
    "schemaId": "namespace/file-base",
    "schemaIdSlug": "namespace_file-base",
    "schemaPath": "/abs/path/to/schema.mjs",
    "scoringProtocol": "v1",
    "scoringInstructions": "Du bist ein Schema-Bewerter. Bewerte unabhaengig basierend nur auf den unten gelieferten Informationen. Antworte als JSON-Array: [ { \"dimension\": \"...\", \"score\": <1.0-5.0>, \"reasoning\": \"...\" }, ... ]. Keine externen Annahmen, keine Kontext-Vermutungen.",
    "prompts": [
        {
            "dimension": "whenToUse",
            "prompt": "Rate the clarity and specificity of the following Schema description on a scale 1.0-5.0. Schema description: \"...\""
        },
        {
            "dimension": "parameters",
            "prompt": "Rate how well the parameter descriptions in the following schema enable an LLM to call the tools correctly on a scale 1.0-5.0. Schema: {...}"
        }
    ]
}

scores.json — Written by Grader

Section titled “scores.json — Written by Grader”

Read by flowmcp grading non-deterministic <schema> --consume-scores <path>.

{
    "schemaIdSlug": "namespace_file-base",
    "scoringProtocol": "v1",
    "creator": {
        "skill": "grade-score-single",
        "skillVersion": "1.0.0",
        "session": "claude-code-2026-05-18-03-15"
    },
    "harness": {
        "name": "claude-code",
        "version": "1.0.0",
        "model": "claude-opus-4-7",
        "modelContext": "1M"
    },
    "timestamp": "2026-05-18T03:15:00Z",
    "scores": [
        {
            "dimension": "whenToUse",
            "score": 4.0,
            "reasoning": "Description is clear and specific..."
        },
        {
            "dimension": "parameters",
            "score": 3.5,
            "reasoning": "Tools have descriptive names..."
        }
    ]
}

Grade Report — Written by CLI

Section titled “Grade Report — Written by CLI”

flowmcp grading non-deterministic <schema> --consume-scores <path> writes the final report:

{
    "schemaId": "namespace/file-base",
    "schemaIdSlug": "namespace_file-base",
    "schemaPath": "/abs/path/to/schema.mjs",
    "schemaHash": "sha256:abc...",
    "date": "2026-05-18",
    "grade": "B",
    "score": 3.75,
    "scoringProtocol": "v1",
    "creator": { ... },
    "harness": { ... },
    "timestamps": {
        "startedAt": "2026-05-18T03:00:00Z",
        "scoredAt": "2026-05-18T03:15:00Z",
        "gradedAt": "2026-05-18T03:16:00Z",
        "reportedAt": "2026-05-18T03:16:00Z"
    },
    "dimensions": [ ... ],
    "validationPassed": true,
    "validationErrors": []
}

Reproducibility Guarantees

Section titled “Reproducibility Guarantees”

The protocol provides statistical reproducibility (not bitwise):

GuaranteeHow
Same schemaschemaHash (sha256) in report
Same prompt templatescoringProtocol: "v1" in scores + report
Same LLMharness.model in scores + report
Same scoring contextGrader MUST use isolated context (no session-state)
Same output formatJSON-Schema for scores

Expectation: same inputs → same score within ±0.2 on the 1.0-5.0 scale.

Schema ID Slug Convention

Section titled “Schema ID Slug Convention”

Schema IDs follow the <namespace>/<file-base> pattern (e.g. mudab/marine-data).

For use in file paths, the / is replaced with _:

mudab/marine-data → mudab_marine-data
coingecko-com/search-ohlc → coingecko-com_search-ohlc

Reference implementation: flowmcp-core/v4/GradeReporter._formatSuggestedFileName.

Grader Implementation Requirements

Section titled “Grader Implementation Requirements”

A Grader MUST:

  1. Read prompts.json, parse scoringProtocol field
  2. Reject unknown protocol versions
  3. Score each prompt independently in an isolated context (no shared state between prompts of the same schema)
  4. Return scores in the same order as input prompts (or include dimension field for matching)
  5. Set creator, harness, timestamp metadata
  6. Write scores.json atomically (write-temp + rename)

A Grader MUST NOT:

  • Use external context beyond what’s in prompts.json
  • Use tools or external lookups
  • Embed scores from previous sessions
  • Modify the scoringProtocol value

Reference Implementations

Section titled “Reference Implementations”
ImplementationTypeLocation
grade-score-singleWorkbench Skillprojects/memo-init/repos/core/skills/grade/grade-score-single/SKILL.md
grade-score-batchWorkbench Skillprojects/memo-init/repos/core/skills/grade/grade-score-batch/SKILL.md
flowmcp grading non-deterministicCLI Producer/Consumerflowmcp-cli/src/task/FlowMcpCli.mjs — gradingDeterministic() / gradingNonDeterministic()
GradeReporterCore Moduleflowmcp-core/src/v4/task/GradeReporter.mjs

References

Section titled “References”
Section titled “Related”