NagaYu/claimcheck-eval

ClaimCheck Evaluation Set

158 hand-checked cases for claim-level groundedness verification: given an answer and the context that was supplied to the model, is each specific claim in that answer supported, derived, approximate, unsupported or contradicted?

Bilingual (87 English / 71 Japanese). Every case was run through ClaimCheck v1.1.0, and the dataset records what the tool actually returned, not just what it should have returned.

Cases	158
Languages	English 87 · Japanese 71
Claim types	NUMERIC 59 · DATE 37 · QUOTE 27 · ENTITY 22 · URL 13
Format	JSONL, one case per line

Related: Space · rule pack

---

Why three judgement columns

Most verifier benchmarks carry one label, which quietly conflates two different questions. This set separates them.

column	question it answers
expected_status	What do the documented rules say should happen?
claimcheck_status	What did the implementation actually return?
human_verdict	Does that outcome actually help a user?

agrees compares the first two. human_verdict is independent of both: a case can be perfectly spec-conformant and still a false positive. That is the whole point. A verifier that scores 100% against its own specification can still be useless if the specification flags things humans accept.

human_verdict:  acceptable 133 | false_positive 22 | false_negative 3

13.9% of this set are false positives — cases where ClaimCheck flags an answer a careful human would accept. That number is published deliberately. A verifier's false-positive rate is the number that decides whether anyone keeps it switched on, and hiding it helps nobody.

---

Schema

field	type	description
case_id	string	cc-000 … cc-157
source	string	authored (78, hand-written) or generated (80, model-generated then hand-verified)
language	string	en / ja
claim_type	string	NUMERIC / DATE / QUOTE / ENTITY / URL
category	string	grounded, derived, contradicted, hallucinated, paraphrase, coverage, adversarial, edge_case, known_false_positive, known_false_negative
context	string	The source text supplied to the model
answer	string	The model output under test
focus_value	string	The exact substring of answer under test (empty for coverage cases)
expected_status	string	Spec-conformant label, incl. not_extracted
claimcheck_status	string	What ClaimCheck v1.1.0 returned
agrees	bool	expected_status == claimcheck_status
human_verdict	string	acceptable / false_positive / false_negative
evidence	string?	The tool's justification (e.g. the derivation formula)
nearest_context_value	string?	For contradictions, the near-but-different context value
similarity	float?	For fuzzy matches
verdict	string	Answer-level gate verdict: pass / annotate / retry / block
grounding_score	float	Of the claims checked, the share that held up
coverage	float	Share of sentences that produced any checkable claim
claims_total, sentences_unverified	int
latency_ms	float	Verification time for this case
rationale	string	One sentence on why the label is what it is

Statuses

status	meaning
supported	Appears literally in the context after normalisation
derived	Computable from context numbers by 2–3 term arithmetic; the formula is in evidence
approximate	A near match exists — paraphrase, finer-grained date, related URL
unsupported	Nothing in the context backs it
contradicted	The context holds a near but different value
not_extracted	No claim was produced at all — relative dates, list markers, vague prose. This is a valid outcome, and it is what lowers coverage rather than being scored

---

Results for ClaimCheck v1.1.0

spec conformance                158/158  (100%)
answer-level recall              56/56   (100%)   cases that should be flagged, and were
clean answers passing            44/46   (96%)    the other 2 were annotated, never blocked
known false positives            22      (13.9%)
known false negatives            3       (1.9%)
latency                          p50 0.17ms · p95 0.39ms · max 0.96ms

Conformance is 100% by construction — the labels were reconciled against the documented rules, and seven implementation bugs found during that reconciliation were fixed rather than labelled around. Conformance is therefore not a quality score. The numbers that carry information are the last three rows.

Answer-level recall vs claim-level status

Every one of the 56 cases that should be flagged produced a non-pass verdict, including the three claim-level false negatives. This is the layered design working: when a quote's similarity is too high to call it contradicted, the NUMERIC or ENTITY claim inside that quote still fires.

cc-093  「本件は民法第719条に基づく不法行為に当たる」   (context: 第709条)
        QUOTE  → approximate      (0.97 similarity: reads as a paraphrase)
        ENTITY → contradicted     ← the layer that catches it
        answer verdict → retry

Read the claims inside a quote, not only the quote's own status.

---

Known failure modes

The most important one: Japanese semantic reversal

All three false negatives are the same shape. A one-character change that reverses meaning keeps character similarity above 0.9, so the quote reads as a faithful paraphrase:

context	answer	similarity	status
国内売上高は前年同期比で7%増加した	国内売上高は前年同期比で7%減少した	0.97	approximate ❌
提案は全会一致で承認された	提案は全会一致で否決された	0.92	approximate ❌
Subscription revenue grew 8%	Subscription revenue declined 8%	0.97	approximate ❌

Character-level matching cannot detect negation or antonym substitution. If your domain turns on approve/reject, increase/decrease or grant/deny, do not rely on quote status alone.

False positives (22 cases, 13.9%)

Distribution: QUOTE 6 · ENTITY 6 · NUMERIC 5 · DATE 4 · URL 1.

• Faithful paraphrase — meaning preserved, wording replaced, similarity falls below threshold. The largest single group.
• Generic noun phrases — "Machine Learning" flagged as an unsupported entity.
• Acronym expansion — context "World Health Organization", answer "WHO".
• Rounding — context 12,345, answer roughly 12,000: 2.8% off, inside the near-miss window. Raise numeric_tolerance rather than lowering contradiction_rel.
• Fiscal vs calendar year — FY2024 and 2024 treated as the same period.
• Date arithmetic — 昭和48年 + 創業50周年 = 2023 is correct, but derivation is arithmetic over numbers only, so the date is not derived.
• Identifier variants — ISO 9001:2015 where the source says ISO 9001. The deliberate cost of catching one-digit-off fabrications.
• URL query parameters — a tracking parameter breaks exact matching.

ENTITY is the noisiest check. Disable it first if you see too much noise.

---

Usage

from datasets import load_dataset

ds = load_dataset("NagaYu/claimcheck-eval", split="test")

# Conformance of your own verifier
for row in ds:
    predicted = my_verifier(row["answer"], row["context"], row["focus_value"])
    print(predicted == row["expected_status"], row["rationale"])

# The rows that actually matter: where flagging hurts the user
fps = ds.filter(lambda r: r["human_verdict"] == "false_positive")

Without datasets:

import json
rows = [json.loads(l) for l in open("data/eval.jsonl", encoding="utf-8")]

Using it as a regression suite

This set doubles as ClaimCheck's own regression suite. It found seven bugs in v1.0.0 that unit tests had missed:

1. \b never fires between a CJK character and a digit, so ISO dates inside Japanese text were invisible (契約は2024-01-05に).
2. U+30FC, the katakana prolonged sound mark, was folded to a hyphen, turning データセンター into デ-タセンタ- and breaking Japanese matching wholesale.
3. An exact sign flip was excused by a coincidental 3-term derivation — a loss reported as a profit returned verdict: pass.
4. Leading whitespace was absorbed into numeric matches ("\n2"), breaking list-marker suppression and span accuracy.
5. ISO currency codes were claimed as entities, so USD 45,000 lost its number.
6. A bare year in the answer was never compared against dates in the context.
7. Parenthesised accounting negatives (1,200) were read as positive.

Cases 1, 2, 3 and 5 all affect Japanese or financial text specifically, and none of them surfaced in English-only smoke tests. That is the argument for a bilingual evaluation set.

---

Construction

• 78 authored cases written against the specification, covering each claim type, both languages, and deliberate known-limitation rows.
• 80 generated cases produced by five parallel Claude Opus 5 agents, one per claim-type group, then hand-verified. Every focus_value was checked to be an exact substring of its answer; labels that were wrong about the documented rules were corrected through an auditable override table with a written reason for each.
• Both sets were then run through the live implementation to populate claimcheck_status.

All entities are fictional (example.com, Acme Corp, 架空商事). No real personal data, and no real organisation is described as doing anything.

Limitations

• 158 cases is small. Treat the false-positive rate as indicative, not precise.
• Domain skew toward business, finance and technical documentation.
• Single-annotator labels. No inter-annotator agreement was measured.
• Tuned against one implementation. expected_status follows ClaimCheck's rules; another verifier may reasonably disagree, particularly on the approximate / contradicted boundary.
• Generated cases were reviewed by the same author who wrote the authored cases.

Citation

@misc{claimcheck_eval_2026,
  title  = {ClaimCheck Evaluation Set: claim-level groundedness verification in English and Japanese},
  author = {NagaYu},
  year   = {2026},
  url    = {https://huggingface.co/datasets/NagaYu/claimcheck-eval}
}

Apache-2.0.