sileod/cordis-bench

CordisBench

CordisBench tests whether language models can reason about the consequences of component lifecycle changes in dynamic agent harnesses. Each record contains an exact, programmatically generated oracle. Set-valued tasks use Jaccard similarity, sequence prediction uses per-observable accuracy, and executable reconfiguration is checked by running the proposed lifecycle operations.

This repository packages the frozen V2.0.1 release from sileod/cordis-bench.

Splits

Split	Rows	Purpose
core	432	Small semantic instances and interface checks
challenge	624	The six-size primary scaling evaluation
diagnostic	144	Secondary outcome-count diagnostics

challenge_diamond is reserved for a future, deeper public challenge. It will use the same schema and verifier, but is not part of V2.0.1 and has not been evaluated in the paper.

Prediction format

Submit one JSON object per line. response may contain the model's raw answer; the official parser extracts the requested value. Advanced runners may instead provide a canonical parsed_answer.

{"id": "e990e7541786d20e", "response": "[\"A\",\"B\",\"D\"]"}

Every ID in the selected split must occur exactly once. Unknown and duplicate IDs are rejected.

Verification

Install the verifier dependency and score a file:

python -m pip install -r requirements.txt
npm ci --prefix native/cordis
python verify.py predictions.jsonl --split challenge --output report.json

The command prints the headline metrics and writes the complete task-level report when --output is supplied. Reconfiguration answers are executed by the pinned Cordis runner, so Node.js and the npm ci setup step are required when the selected split contains those tasks. Other tasks need only Python.

To rebuild the Hub files deterministically from the benchmark checkout:

python build_release.py --source ../cordis-bench/data/generated/v2.0.1-release.jsonl

Pass --diamond PATH once a future challenge_diamond JSONL has been frozen. The builder validates IDs and writes counts and SHA-256 checksums to checksums.json.

Metrics

• Localization, guaranteed conditions, and reachable conditions: Jaccard
• Schedule prediction: accuracy per predicted observable
• Reconfiguration: executed success rate
• Whole-answer exact match and parse rate: strict diagnostics

The verifier also retains detailed rows so alternative summaries can be audited without rerunning model inference.