ChainCheck

A benchmark kit for evidence-sufficiency scorers in RAG that separates two things an ordinary benchmark mixes up: reacting to a broken reasoning chain, and reacting to any edit in the passages. 3,786 items over MuSiQue and 2WikiMultiHopQA, with a standalone evaluator.

An ordinary sufficiency benchmark asks whether a scorer ranks untouched passages above broken ones. Near-perfect scores on that question are common, and they can come from a scorer that flags every edited context. ChainCheck adds a matched control, so a scorer that tracks the chain and a scorer that tracks edits no longer look the same.

Three cells

Cell	Chain	Edit
A	intact	none
B	intact	consistent bridge-entity swap across all passages
D	broken (substitution in the later hop only)	yes

• Nominal = AUC(A > D): what an ordinary benchmark reports.
• Chain effect CE = AUC(B > D) − 0.5: edit held fixed, chain varied.
• Edit effect EE = AUC(A > B) − 0.5: chain held fixed, edit varied.
• Chain selectivity Σ = CE − |EE|. A 95% lower bound above zero means the scorer reacts more to chain breaks than to edits.

These are matched contrasts, not an additive decomposition (Nominal is not CE + EE), and the fourth cell (broken chain, no edit) is not constructed.

Getting started

from datasets import load_dataset
ds = load_dataset("ThakiCloud/ChainCheck", "twowiki_replication", split="train")
print(ds[0]["cell"], ds[0]["query"], len(ds[0]["passages"]))

Running the evaluation

Score every item with your system (higher = more sufficient), write one {"item_id": ..., "score": float} per line, then:

python chaincheck_eval.py --data twowiki_replication.jsonl.gz --scores my_scores.jsonl --exclude-cb27b

The evaluator needs only numpy. It reports Nominal, CE, EE and Σ with pair-bootstrap intervals (2,000 resamples, seed 0), separately for real and synthetic replacement entities. --exclude-cb27b drops pairs that Qwen3.8-27B answers closed-book, as the paper does.

Files

File	Rows	Source license
musique_test.jsonl.gz	1,716	MuSiQue (Trivedi et al., TACL 2022), CC BY 4.0
musique_confirmation.jsonl.gz	678	MuSiQue (Trivedi et al., TACL 2022), CC BY 4.0
twowiki_replication.jsonl.gz	1,392	2WikiMultiHopQA (Ho et al., COLING 2020), Apache-2.0

Each row is one (question, passage set) item: item_id, pair_id, variant (real, or fict for a synthetic replacement entity), cell (A/B/D), query, passages, answer, bridge, replacement, chain, edit, earlier_support, later_support, label_sufficient, cb_27b_correct and source. Only quadruples that pass the Q1 to Q9 validator are included.

Baseline (Qwen3.8-27B zero-shot judge)

File	Variant	Pairs	Nominal	Chain effect	Edit effect	Σ [95% CI]
MuSiQue test	real	324	0.895	+0.309	+0.108	+0.201 [+0.123, +0.278]
MuSiQue test	fict	224	0.906	+0.397	-0.009	+0.388 [+0.312, +0.420]
MuSiQue confirmation	real	122	0.943	+0.344	+0.164	+0.180 [+0.066, +0.295]
MuSiQue confirmation	fict	98	0.908	+0.449	-0.031	+0.418 [+0.316, +0.469]
2Wiki replication	real	295	0.905	+0.320	+0.192	+0.129 [+0.054, +0.200]
2Wiki replication	fict	141	0.929	+0.443	+0.046	+0.397 [+0.305, +0.454]

Scored with the same pipeline as the released judges. The paper's own run of this judge ranks items essentially identically; its Σ values differ from this table within the reported intervals because A and B score almost the same and bf16 differences flip some comparisons.

What not to trust

• Nominal is near its ceiling for almost every system. Do not rank systems by it alone; that is the failure this kit exists to expose.
• A and B differ by one entity and often score almost the same, so EE is sensitive to small numerical differences in the scorer. Compare systems scored on the same hardware and precision, and read Σ with its interval.
• MuSiQue confirmation has 122 real pairs after filtering; its intervals are wide.
• Edits are entity substitutions on two-hop chains over Wikipedia. Numerical errors, temporal staleness and longer chains are not covered.

Provenance

Labels follow mechanically from source provenance and the validator; no human or LLM labels. musique_test and musique_confirmation derive from MuSiQue (Trivedi et al., TACL 2022); twowiki_replication derives from 2WikiMultiHopQA (Ho et al., COLING 2020), compositional and inference question types. The confirmation slice and the 2Wiki replication were built after the protocol was frozen.

License

Each file keeps its source license: MuSiQue-derived files are CC BY 4.0, the 2WikiMultiHopQA-derived file is Apache-2.0, and passage text originates from Wikipedia (CC BY-SA). The evaluator script is Apache-2.0. Cite the original datasets alongside this one.

Models

Judges trained with these controls (4B, 9B, 27B): ChainCheck collection.

Paper

ChainCheck: When Near-Perfect Evidence-Sufficiency Scores Fail to Distinguish Chain-Sensitive Systems — preprint forthcoming