benchmark-core

AI Benchmark Index · npm · API · MCP

benchmark-core is an open-source TypeScript engine for reproducible AI model benchmarks. One definition produces an SDK, command-line interface (CLI), Model Context Protocol (MCP) server, and HTTP API.

It measures whatever your tasks and deterministic grader define. The included example measures 20 arithmetic answers and one tool-use episode with a transient order-service fault.

npx benchmark-core run --model offline --provider mock --json

{"model":"mock:1","mean":0.5238,"confidenceInterval95":{"low":0.3333,"high":0.7143},"totals":{"tasks":21,"passed":11,"failed":0}}

The example data comes from a committed seeded generator. Seed 20261015 reproduces every public task.

Install

benchmark-core requires Node.js 20 or newer.

npm install benchmark-core

The package has one runtime dependency: the official MCP TypeScript SDK.

Define a benchmark

import { defineBenchmark, exactMatch, run, mockAdapter } from "benchmark-core";

const benchmark = defineBenchmark({
  id: "benchmark-colors",
  name: "Color recall",
  tagline: "Checks exact color recall.",
  tags: ["science"],
  category: "memory",
  whatItMeasures: "Exact recall of a named color.",
  version: "0.1.0",
  status: "alpha",
  tasks: [{ id: "sky", kind: "single", prompt: "Name the sky color.", expected: "blue" }],
  grade: (task, submission) => exactMatch(task.expected, submission.answer ?? ""),
  provenance: [{
    name: "Maintainer-written task",
    kind: "human-written",
    url: "https://github.com/example/benchmark-colors",
    licence: "MIT",
    howToGet: "Read data/tasks.jsonl.",
    retrievedAt: "2026-10-15T00:00:00.000Z",
    contaminationNotes: "Public task; use a hidden split for final comparisons.",
    refresh: "Review and add tasks through pull requests."
  }],
  gradingMethod: "Exact string match."
});

const report = await run(benchmark, mockAdapter(), { concurrency: 2, seed: 7 });
console.log(report.mean);

tasks can be replaced by an asynchronous loadTasks() function. Task IDs must be unique, and each task declares kind: "single" | "episode".

How it works

defineBenchmark({ tasks, grade, environment, provenance })
                  │
        ┌─────────┼──────────┐
        │         │          │
       SDK       CLI      MCP server
        │         │          │
        └─────────┴──── HTTP handler
                  │
             RunReport JSON

The runner preserves task order while it uses bounded concurrency. Each row records attempts, latency, score, reason, usage, and errors. Episode rows also contain the transcript, tool log, final state, and abort reason.

The 95% confidence interval uses 2,000 seeded bootstrap resamples. Identical scores and seeds produce identical intervals.

Runtime surfaces

CLI

A benchmark package calls createCli(benchmark, { bin }). The returned main(argv) supports:

• list
• show <task>
• run --model <id> --provider <openrouter|openai|mock> with optional episode limits
• score <results.json>
• card <report.json> [output.svg]
• provenance
• index-entry
• serve [port]
• mcp

Every command accepts --json. Exit code 0 means success, 1 means invalid usage, and 2 means a run or runtime failure.

Model Context Protocol

createMcpServer(benchmark) exposes list_tasks, get_task, score_answer, run_benchmark, and get_provenance. startMcp(benchmark) connects it to standard input and output.

run_benchmark uses OpenRouter and requires OPENROUTER_API_KEY. The other four tools need no network or key.

HTTP API

createHandler(benchmark) returns (request: Request) => Promise<Response>. It has no global state and works with Node.js or Cloudflare Workers.

| Method | Route | Result | | --- | --- | --- | | GET | /health | Package identity and health | | GET | /v1/tasks | All tasks | | GET | /v1/tasks/:id | One task | | POST | /v1/score | Deterministic score for { taskId, submission } | | GET | /v1/provenance | Data sources | | GET | /v1/index-entry | Directory entry |

GET responses include permissive Cross-Origin Resource Sharing (CORS). Request bodies over 1 MiB return a stable REQUEST_TOO_LARGE error.

serve(handler, { port }) hosts the same handler with node:http.

The server listens on 127.0.0.1 by default. Use serve --host 0.0.0.0 only when a container or proxy needs an external bind.

POST /v1/score has no built-in authentication. Put authorization and rate limits in front of public deployments. An opt-in LLM-judged grader makes a paid model call for every score request.

Tool-use episodes

An episode task runs against a seeded, in-process Environment. The environment exposes JSON-schema tool definitions and four deterministic methods:

interface Environment {
  reset(task, seed): { system: string; tools: ToolSpec[] };
  callTool(name, args): { result: unknown; error?: { code: string; message: string } };
  state(): unknown;
  done(): boolean;
}

runEpisode(task, adapter, environment, options) enforces turn, tool-call, and dollar limits. A budget breach returns an episode with aborted: "budget"; the main runner still sends its trace to the grader.

The universal submission is a Trace:

{
  answer?: string;
  toolLog?: ToolLogEntry[];
  finalState?: unknown;
}

replay(environment, task, toolLog, seed) resets the simulator and repeats only recorded calls. Episode graders can compute a result without a model or network call.

The reference OrderLookupEnvironment injects an HTTP 500 on its first seeded lookup. Its good mock policy retries, loads a $35 order, and applies the $50 refund limit.

Model adapters

• mockAdapter({ seed, policies }) works offline. It returns a deterministic mix of expected and wrong answers, then follows benchmark-supplied good(task) or bad(task) episode scripts.
• openrouterAdapter({ model }) reads OPENROUTER_API_KEY.
• openaiCompatAdapter({ baseUrl, model }) reads OPENAI_BASE_URL and OPENAI_API_KEY when options omit them.

Adapters implement complete(prompt, context) for single tasks and chat({ messages, tools }, context) for episodes. API-reported usage and cost flow into each row and report total.

Graders

• exactMatch(expected, answer) compares JSON values exactly.
• regexMatch(pattern, answer) tests string answers.
• numericWithin(expected, tolerance, answer) accepts numbers within an inclusive tolerance.
• jsonSchemaMatch(schema, answer) checks a dependency-free JSON Schema subset.
• judgeWithLLM(adapter, prompt?) creates an opt-in model grader.

jsonSchemaMatch supports type, enum, required, properties, items, additionalProperties, pattern, minimum, and maximum.

The default judge prompt is exported as DEFAULT_JUDGE_PROMPT:

Score the candidate answer from 0 to 1. Return only JSON with numeric score and string reason.

A judge must return {"score": number, "reason": string}. Model judging is off unless benchmark code calls judgeWithLLM.

Reports and metadata

• renderCard(report) returns a 1200×630 SVG share card.
• renderMarkdown(report) returns a Markdown summary and task table.
• indexEntry(benchmark) returns the AI Benchmark Index record.
• provenanceMarkdown(benchmark) renders provenance documentation from code.

Commit the rendered provenance file beside tasks.jsonl. The end-to-end test compares the example file byte for byte.

Utilities and types

mulberry32(seed) provides the small seeded random number generator. bootstrapCI(values, seed?, resamples?) returns percentile bounds.

The package exports these types:

• Task, Trace, TaskScore, and GradeContext
• ToolSpec, ToolCall, ToolLogEntry, ToolResult, and ToolError
• Environment, EnvironmentStart, Episode, and EpisodeOptions
• BenchmarkDefinition, Benchmark, BenchmarkTag, and BenchmarkStatus
• ModelAdapter, ModelAdapterContext, CompletionResult, ChatRequest, ChatResult, ChatMessage, and ModelUsage
• MockPolicies and ScriptedTurn
• RunOptions, RunRow, and RunReport
• Provenance, ProvenanceKind, and IndexEntry
• JsonSchema, JsonSchemaObject, JsonValue, JsonPrimitive, and Answer

Limits

The runner does not hide failed tasks. Exhausted retries become zero-score rows with an error message.

A timeout stops waiting and aborts the provider signal. Provider work can continue in the background if its adapter ignores that signal.

The built-in schema grader implements the documented subset, not the full JSON Schema specification. Cost stays absent when a provider omits it.

The public arithmetic and order tasks test the surfaces, not frontier model ability. Use a private or held-out split for serious comparisons.

Development

npm ci
npm run ci:local

The gate lints, type-checks, builds, and runs one end-to-end script. That script exercises every task, API route, and MCP tool, then writes artifacts/e2e-report.json.

Version 0.1.0 is alpha. Report defects through GitHub issues.

License

MIT © Pooria Arab