# Dataset usage card

Complete sections 1–2 for each release candidate. Complete section 3 for a benchmark task and section 4 only for an optional RSI study. Keep claims proportional to what is actually downloadable and runnable. A standalone resource does not require an agent or an RSI experiment.

## 1. What can I use?

- Dataset / task name, version and scientific scope:
- Status and reviewed scope: scientific resource / benchmark task / teaching integration kit / RSI study package:
- Record count, family or independence unit, and exclusions:
- Access: direct artifact / pinned upstream source / controlled-access instructions:
- Data host, landing page and download or access instructions (the full dataset may remain externally hosted):
- File format, approximate download size and any storage/runtime requirements:
- Immutable version and checksum (or source commit):
- Access last checked: date, actual result and any restrictions (pending until checked):
- Reuse terms and original-source restrictions:
- Citation for the original paper/resource and separate adaptation:

## 2. Can I run a small example?

- Supported environment and installation requirements:
- Smallest permitted example download; identify any included answers:
- Exact commands from a clean extraction or checkout:
- Expected output and what it demonstrates:
- Where outputs are saved; reset/retry and no-overwrite behavior:
- Actual independent replay date, environment and result (leave pending until run):

A passing smoke test establishes the I/O path. Label fixed or synthetic fixtures and keep them distinct from a scientific baseline or measured improvement.

## 3. Benchmark task interface (if applicable)

For a task package, describe how a researcher supplies predictions and checks outcomes. A dataset without this interface can leave this section blank and document its loading example in section 2.

- Allowed solver input schema with one example:
- Output schema with one example; IDs, indexing and parsing rules:
- Loader or adapter entry point; existing framework alternative:
- Checker invocation, missing/invalid outputs, fixed denominator and retries:
- What feedback is available, to whom and when:
- Which files contain answers or sensitive metadata and must stay evaluator-side:
- Actual isolation boundary, or explicitly no reserved evaluation set (shared local files are not protected evaluation):

## 4. Optional RSI study package

Complete this only when contributing an RSI study contract. Ordinary datasets and benchmark tasks can leave it blank.

- Supported editable state and clear solver/improver boundaries:
- How changes are proposed, installed and used in later calls:
- What the package executes now versus what the experimenter must implement:
- Log fields for actual calls, revisions, failures, usage and budgets:
- Independent test access and fixed-improver/matched-budget controls:
- Current evidence: source data / smoke run / scientific baseline / revision trace / independent benefit:

Do not infer RSI from replaying known answers. Revised-improver execution and independent benefit require distinct evidence. An exposed example stays development material; a future evaluation set is not available until it actually exists.