---
license: "mit"
language:
- en
pretty_name: "Schema Mutation"
tags:
- benchmark
- llm-evaluation
- ai-benchmark
- model-evaluation
- evals
- grading
- openrouter
- openai
- mcp
- typescript
- agent-benchmark
- function-calling
- tool-use
- openapi
- schema-migration
task_categories:
- text-generation
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/tasks.jsonl
---

# Schema Mutation

Tool schemas change. Does the model call the current one?

## What it measures

Whether a model follows the current tool schema after a migration — renamed fields, new required fields, changed enums, split endpoints and type changes — instead of the version it memorized.

Category: `tools`. Status: `alpha`.

## How to run

```console
$ npx benchmark-schema-mutation list
$ npx benchmark-schema-mutation run --model <model-id> --provider openrouter
$ npx benchmark-schema-mutation score results.json
```

`--provider mock` runs the whole benchmark offline with no keys. Real models run
through OpenRouter (`OPENROUTER_API_KEY`) or any OpenAI-compatible base URL
(`OPENAI_BASE_URL`, `OPENAI_API_KEY`).

## Data fields

`data/tasks.jsonl` holds 60 tasks, one JSON object per line
(kinds: 60 single).

| field | type |
|---|---|
| `id` | string |
| `kind` | string |
| `prompt` | string |
| `expected` | string |
| `metadata` | object |

## Provenance

# Schema Mutation provenance

This file is generated from the benchmark definition.

## Seeded schema-mutation task generator

- Kind: `synthetic`
- URL: https://github.com/pooriaarab/benchmark-schema-mutation/blob/main/scripts/gen-tasks.mjs
- Licence: `MIT`
- Retrieved: 2026-10-15T00:00:00.000Z
- How to get it: Run `node scripts/gen-tasks.mjs` for the public split (seed 20261015) or `node scripts/gen-tasks.mjs --seed <n> --split hidden` for a held-out split.
- Contamination notes: All 60 public tasks are synthetic mutations of 12 hand-authored base tool schemas. The hidden split uses a different seed, which selects different values, mutation variants, field names and enum targets, and is never committed, so public tasks may enter training data without leaking final evaluation items.
- Refresh: Change the seed, rerun the generator, and review the JSONL diff. Rotate mutation parameters quarterly.


## Grader

Canonical JSON equality on the tool name and arguments after normalisation; exact matches against the pre-migration call are reported as staleSchemaUsed.

## Limitations

This dataset contains only the public task split (60 tasks).
Hidden evaluation splits use a private seed and are deliberately not part of
this dataset. Source kinds: synthetic. These are evaluation prompts
for the `benchmark-schema-mutation` benchmark, not training data; scoring requires the package's
deterministic grader. See PROVENANCE.md for contamination notes.

## Citation

```bibtex
@misc{benchmark_schema_mutation,
  author = {Pooria Arab},
  title = {Schema Mutation task data},
  year = {2026},
  url = {https://github.com/pooriaarab/benchmark-schema-mutation}
}
```

## Links

- npm: https://www.npmjs.com/package/benchmark-schema-mutation
- GitHub: https://github.com/pooriaarab/benchmark-schema-mutation
- Leaderboard: https://aibenchmarkindex.com/benchmarks/benchmark-schema-mutation
- Dataset: https://huggingface.co/datasets/pooria/benchmark-schema-mutation