REO-20 v1.0.0

REO-20 is a frozen synthetic matched-pair benchmark for testing whether a model can decide when a proposed local rule update is authorized by the evidence and then execute the active rule correctly.

Each of the 10 matched pairs contains two closely related items. One decisive piece of evidence changes between the two members. A successful model must both choose the correct rule state and apply that rule correctly on both sides of the pair.

Quick start: reference Colab run

The easiest way to see the complete evaluation workflow is to run the public reference notebook:

REO-20 with Qwen3.5-4B on a Colab T4

In Google Colab:

1. Open the notebook.
2. Select Runtime → Change runtime type → T4 GPU.
3. Run all cells from top to bottom.

The notebook:

1. downloads and validates this REO-20 repository;
2. loads a fixed revision of the official Qwen/Qwen3.5-4B model in FP16;
3. runs all 20 benchmark items independently, without a system message or few-shot examples;
4. saves the raw model responses;
5. runs the scorer included in this repository;
6. downloads a results ZIP containing predictions, scores, runtime logs, and reproducibility metadata.

The notebook is a concrete reference implementation, not a requirement to use that specific model.

Repository map

• data/reo20_items.frozen.jsonl: the actual benchmark data, containing 10 pairs / 20 items.
• README.md: overview, quick start, and repository map.
• EVALUATION_PROTOCOL.md: the model-facing evaluation procedure.
• scripts/score_reo20.py: command-line scorer for model predictions.
• scripts/validate_release.py: repository integrity and release validator.
• tests/gold_predictions.jsonl: a known-correct example of the prediction format.
• tests/: parser and scorer tests.
• controls/direct_controls.jsonl: diagnostic items that test direct execution of the released rules.
• src/: the frozen parser, scorer, and prompt wrapper implementation.
• schemas/: data schemas.
• spec/: the frozen scorer contract and task wrapper specification.

The files needed to understand the basic workflow are data/reo20_items.frozen.jsonl, EVALUATION_PROTOCOL.md, and scripts/score_reo20.py. The remaining files support validation, diagnostics, and reproducibility.

The Hub Dataset Viewer is explicitly configured to load only data/reo20_items.frozen.jsonl. Python files are ordinary reproducibility files, not dataset-loading scripts. All public file names, documentation, code comments, diagnostics, and data are English-only.

Running another model

A different model, API, or inference framework can be used as long as the same evaluation contract is preserved:

1. Read each record from data/reo20_items.frozen.jsonl.
2. Send the record's rendered_prompt unchanged as one independent user message.
3. Do not add a system message, few-shot demonstrations, or explanatory text.
4. Preserve the model's raw response exactly.
5. Write one JSON Lines record per item using member_id as item_id.

Prediction format:

{"item_id":"<member_id>","response":"<raw model response>"}

Then run:

python scripts/score_reo20.py --predictions predictions.jsonl --output scores.json

For a reproducible report, also record the model repository and revision, tokenizer or chat template, inference framework, numerical precision or quantization, hardware, and generation settings.

Output contract

Return exactly:

DECISION: KEEP_CURRENT or APPLY_UPDATE
ACTIVE_RULE: R0 or R1
RESULT: <result>

The scorer evaluates the update decision, selected active rule, executed result, formatting, and matched-pair behavior.

Validation

Run the repository validator and tests with:

python scripts/validate_release.py
python -m unittest discover -s tests -v

The bundled src/parse.py, src/score.py, and src/wrapper.py are copied from the frozen REO-20 v1.0.0 artifact. The public CLI is only an adapter around those frozen components.

Construction and selection

The set was selected using structural quality and coverage constraints. Model results and PCE/EPE results were not used for selection. Reserve items, raw model outputs, review ledgers, and internal registries are not distributed.

License

All released files are dedicated to the public domain under CC0 1.0 Universal. Citation is appreciated but not required.

Lineage

• Frozen artifact ID: reo20-frozen-release:4e05cb8c-afdb-4849-b918-002400c0a6d3
• Frozen source ZIP SHA-256: 0b9cca4d451b0e9c9f041a167954d13d1c13bbfe47c2c4a4f2f89ad71e4d4854
• Frozen main data SHA-256: 8192febe8b60e498ad7f145a925d4fb9c06e3d8ddf424525c3cc3d7047be6723