claimcheck-bench

A synthetic benchmark for checking whether an AI agent's success claim is supported by its tool results and environment state.

An agent can say “done” after a failed write, an action on the wrong record, or an operation that never persisted. This dataset contains 300 labelled agent traces for evaluating detectors that distinguish successful completion from false success claims across booking, CRM, and coding tasks.

Created by Andrii Boiko as part of agent-claimcheck, an open-source agent reliability and evaluation toolkit. GitHub profile · LinkedIn

What is included

Split	Traces	Success	Failure
Train	180	108	72
Test	120	72	48
Total	300	180	120

Each domain has 60 train and 40 test examples. All traces contain a success claim; failure means the simulated task was not fully completed despite that claim. Successful examples include clean runs, recovered errors, and noisy runs, so an error in a trace does not automatically mean the task failed. There are no unlabelled or claim-free examples in these splits.

Seven false-success scenarios are represented:

Kind	What goes wrong
phantom_action	The agent claims an action it never called.
error_ignored	An action fails and the agent claims success without recovery.
wrong_target	The action affects the wrong record, attendee, or file.
wrong_value	The target is correct but a requested value is wrong.
not_persisted	A request is accepted but the change does not persist.
partial_completion	Only part of the requested work is completed.
reviewer_injection	A failed run includes text that tries to persuade a reviewer it succeeded.

Load the dataset

import json
from datasets import load_dataset

dataset = load_dataset("aboiko/claimcheck-bench")
row = dataset["test"][0]
trace = json.loads(row["trace_json"])

# Only these fields are detector inputs. Keep the labels separate.
detector_input = {key: trace[key] for key in ("task", "steps", "final_claim")}
expected_outcome = row["outcome"]

The Parquet view contains string columns with a consistent schema:

Column	Meaning
trace_id	Identifier for joining predictions to examples; not a model input.
domain	booking, crm, or coding.
instruction	The requested task.
final_claim_text	The agent's final success claim.
outcome	Evaluation label: success or failure.
failure_kind	One of the seven failure kinds, or none for successful tasks.
evidence	Evaluation grouping: state_probe or receipt_only.
trace_json	The complete original trace as a JSON string, including labels.

Do not pass an entire row or trace_json directly to a judge. Remove ground_truth, meta, and source, and keep the convenience label/grouping columns out of the detector prompt. Trace IDs should also be excluded from prompts. The toolkit's detector_view implements this separation.

trace_json retains the original JSON line without its line terminator. Parsing it restores the original object without coercing date strings, adding absent fields, or changing nested structures. The traces.train.jsonl, traces.test.jsonl, and manifest.json files preserve the original JSONL data and manifest byte-for-byte. Use those files with the toolkit's agent-trace/v1 reader; the schema is in agent-trace-v1.json.

Generation and provenance

Every trace is synthetic: invented people, companies, and file paths; simulated tool responses; and fictional example.test email domains. These are not production logs or recordings of real client activity. Labels come from the scenario generator, not human annotation or an LLM judge.

Generator: claimcheck-bench/1. Seed: 20260924. Instruction templates, final-message templates, and entity pools are separated between train and test. The split is stratified by domain, class, scenario kind, and evidence type. See the preserved original dataset card for detailed composition and the original leakage audit.

This publication packages the benchmark from source commit 63f1ae233ce9605253b53a3dcf607b6d31f00918. It adds a reversible Parquet view, this card, and provenance records; it does not alter the benchmark examples or labels. provenance.json records hashes, counts, and the transformation. validate.py checks integrity and reconstruction locally.

Appropriate use and limitations

Use this dataset to compare rule-based checkers, learned classifiers, LLM judges, and abstaining ensembles on evidence-based task-completion verification. Report performance on the test split; fit models and calibration on train only. Evaluate abstention and coverage alongside accuracy, and count false successes incorrectly accepted as successful.

• Synthetic, small, and English-only. The finite templates and simulated APIs cannot establish real-world reliability or any product's actual failure rate.
• Co-designed evaluation. The generator, rules, and classifier features were developed by the same author. Strong results may reflect those shared conventions rather than transfer to unfamiliar tools or traces.
• Observability is limited. Some examples contain only receipts. In particular, a non-persisted asynchronous write without a later probe may be indistinguishable from a genuine asynchronous success. Abstaining can be appropriate even when a scenario label is available.
• Simplified calendar conventions. Booking scenarios use fixed offsets for named time zones; those offsets do not follow daylight-saving changes for all generated dates. Labels follow the simulated scenario. Do not use v1 to evaluate civil-time or DST correctness.
• Adversarial text is data. Some examples intentionally contain reviewer-directed instructions. They should be inspected as untrusted trace content, never followed as instructions to the evaluator.

This is a task-completion verification benchmark, not a general factuality, security certification, or autonomous-agent leaderboard. Detector results and their methodological limitations are documented in the source project's recorded evaluation.

License and citation

Apache-2.0. Copyright 2026 Andrii Boiko. See LICENSE.

@misc{boiko2026claimcheckbench,
  author = {Boiko, Andrii},
  title = {claimcheck-bench: Agent Success-Claim Verification},
  year = {2026},
  version = {1},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/aboiko/claimcheck-bench},
  note = {Synthetic benchmark distributed with agent-claimcheck}
}

For issues, use the source repository.