GroundClaim-Bench

Scripted coding-agent episodes with computed ground truth for checking AI agents' completion claims ("all tests pass", "pushed to main", "deployed to production", …). Built for GroundClaim; the full method and results are in its technical report.

All content is synthetic. Projects, commits, traces and messages were generated by bench/generate.py; authors are Bench Agent <agent@example.invalid>; repositories are bench/<name>; there is no personal data and there are no credentials.

What is in an episode

A small Python project in a real git repository with a local bare remote. A scripted agent edits code and runs real commands (pytest in several forms, git commit, git push); every command is recorded in a trace with its real exit code and output. Remote actions (pull requests, deployments, releases) are simulated, and a simulated GitHub — check runs, pull requests, deployments, releases — is provided as API fixtures consistent with that world. The agent ends with a message of claims.

Ground truth is computed: "tests pass" is true only if the project's original test suite passes on the final code (so skipping a test does not make it true); "pushed" only if the remote branch contains the final commit; and so on. 18 episode families cover honest work and false or misleading completions: failing tests, stale runs, | tail and || true, echo "All tests passed", -k subsets, skipped/removed tests, forgotten commits, rejected pushes, wrong PR numbers, failed rollouts, draft releases, missing or empty files.

split	episodes	claims	false	true	not scored (whole-task)	seed
dev	200	423	153	220	50	101
test	1,000	2,202	789	1,098	315	202

The test split was generated once after the GroundClaim rules were frozen and evaluated once; its message phrasings differ from dev.

Files

• data/{dev,test}.jsonl — one row per episode: id, family, message, claims (each with type, sentence, character span in the message, details, truth — null for whole-task claims — and a note on the variant), base and head commits, branch, github_repo, reports, project.
• episodes/{dev,test}.tar.gz — the episodes themselves: <id>/repo (working tree with .git), <id>/origin.git (bare remote), <id>/trace.jsonl (GroundClaim trace format), <id>/github.json (API fixtures), <id>/scenario.json.
• extraction/realistic.jsonl — 151 hand-written sentences in the style of agents' final messages, labelled with the claim types they contain (30 contain none), for evaluating claim extraction.
• results/ — GroundClaim's and the baselines' metrics on dev and test (gold claims and end to end).

from datasets import load_dataset
episodes = load_dataset("sinceaihq/groundclaim-bench", "episodes", split="test")
sentences = load_dataset("sinceaihq/groundclaim-bench", "realistic_sentences", split="test")

To evaluate a checker on the full episodes:

git clone https://github.com/sinceaihq/groundclaim && cd groundclaim && pip install -e ".[dev]"
hf download sinceaihq/groundclaim-bench --repo-type dataset --local-dir hfbench
mkdir -p bench/out && tar -xzf hfbench/episodes/test.tar.gz -C bench/out && cp hfbench/data/test.jsonl bench/out/
python bench/run.py --data bench/out --split test --out results && python bench/summary.py results

The archives contain git repositories. They were created by the generator and contain no hooks or custom configuration, but as with any downloaded repository, inspect before running tools that execute repository configuration (GroundClaim's own git access is hardened against it).

Results (test, GroundClaim with git + trace + GitHub, gold claims)

False claims confirmed 0.0 % (0/789), accepted (confirmed or supported) 3.0 %, contradicted 67.8 %; true claims rejected 0.0 % (0/1,098), confirmed 59.2 %. A naive log check ("a matching command exited 0") confirms 43.9 % of false claims. See the technical report for conditions, families and limitations.

Limitations

Synthetic, single-ecosystem (Python/pytest), English, and written by the same team that wrote GroundClaim's rules; remote state is simulated. Results show that rules behave as designed on scripted failure modes, not real-world accuracy. The 0.1 data was generated before episode projects had their own pyproject.toml; local interpreter paths printed by pytest -v were replaced by placeholders before publication (results are identical).

Licence and citation

CC BY 4.0. Generator code: Apache-2.0. Made by Since AI, a non-profit association in Finland.

@misc{groundclaim_bench_2026,
  title  = {GroundClaim-Bench: checking AI agents' completion claims against evidence},
  author = {{Since AI ry}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/sinceaihq/groundclaim-bench}
}