Datasets:

stanjsx
/

NeuroClaimBench

File	Records	Contents
`cases.jsonl`	544	Scientific questions, analysis plans, source identities, and record links
`references.jsonl`	544	Reference labels, evidence basis, agreement, and scoring eligibility
`tasks.jsonl`	489	Executable plans and identities of their prepared input cohorts
`outcomes.jsonl`	489	Recorded CONFIRM gate results and execution status

Cases

benchmark_case_id identifies the case; claim_uid identifies its claim.
semantic_cluster_id groups related claims for clustered analyses.
benchmark_track is scientific, synthetic_stress, or external_transfer.
target_family, modality, and question describe the scientific target.
contract contains the analysis plan. reference_id links to its reference; task_ids lists its executable tasks and is empty for excluded cases.
provenance, aliases, and alignment fields identify source records and the match between each question and its analysis plan.

References

disposition is confirm (reference supports the claim), abstain (reference withholds support), or unresolved (insufficient evidence).
basis is literature, constructed_control, or unresolved.
strength separates literature references into strict and provisional groups. The provisional group contains additional labels admitted under two-model direct-evidence rules. Neuroimaging experts finalized the literature labels. Controls use constructed; unresolved references use evidence_gap.
score_eligible selects references for agreement metrics.
evidence_ids, derivation_rule, agreeing_models, agreement_pattern, and vote_counts record the evidence and rule supporting each label.

confirm in a reference is a literature or control label. CONFIRM's reporting decision is computed independently from the gate results in outcomes.jsonl.

Tasks

task_id identifies the task; benchmark_case_id links to its case.
dataset_id, discovery_cohort, and replication_cohorts identify inputs.
contract is the executable analysis plan; contract_sha256 identifies it.
partition_hashes identifies the required prepared files for tasks without a generator specification. partition_paths records their source paths.
generator_spec describes control construction; generator_artifact_hashes identifies the associated construction artifacts.
evidence_role and evidence_freshness describe the evidence partition.
code_sha, schema_version, and gate_policy_version identify the implementation and policy used to define the task.

Outcomes

task_id joins the outcome to its task; status records execution status.
gate_verdict contains the gate checks, diagnostic label, and rationale.
confirm_outcome and raw_final_label record the Full-standard result.
error describes a failed execution when present.
cohort_content_hashes identifies analyzed input files. generated_cohort_hashes identifies generated files required by tasks with a generator_spec.
task_fingerprint and detailed_result_sha256 identify the task and its detailed result artifact; the package distributes the compact outcome.

Results and supporting records

File	Contents
`benchmark_splits.json`	Track memberships, adjudication candidates, and excluded-case IDs
`benchmark_results.csv`	Case-level references, outcomes, and evaluation coverage
`benchmark_summary.json`	Metrics by track, target, reference basis, and strength
`benchmark_strata_summary.csv`	Counts, rates, Wilson intervals, and cluster-bootstrap intervals by evaluation group
`target_reference_summary.csv`	The same summaries grouped by target family
`cluster_bootstrap_sensitivity.csv`	Wilson and semantic-cluster-bootstrap interval comparison
`reference_agreement_audit.csv`	Counts by evidence basis, strength, and model-agreement pattern
`unresolved_case_summary.csv`	Unresolved cases, agreement patterns, and recorded outcomes
`alignment_policy.json`	Rules for question-plan alignment and execution eligibility
`analysis_audit.json`	Alignment, evidence, execution-coverage, and result-consistency checks
`analysis_manifest.json`	Analysis input/output hashes and bootstrap settings
`v2_to_v2.1_crosswalk.csv`	Source-to-benchmark ID and contract-hash mappings
`feedback_parent_crosswalk.csv`	Links from follow-up experiment parent claims to benchmark cases
`feedback_reference_summary.csv`	Follow-up parent support grouped by reference category and search budget
`feedback_crosswalk_manifest.json`	Hashes and matching counts for the follow-up parent links

Source paths and artifact hashes in these records describe the benchmark's scientific provenance. The files needed for scoring are included in this package. Task execution uses separately obtained prepared cohort files.

Package files

File	Contents
`README.md`	Installation and quickstart
`USAGE.md`	Loading, metrics, scoring, and task execution
`DATA_ACCESS.md`	Source cohorts, table format, and exact-input requirements
`DATA_DICTIONARY.md`	File and field definitions
`benchmark_manifest.json`	Record filenames, hashes, schema version, counts, and source identities
`RELEASE_MANIFEST.json`	Benchmark identity, record counts, and distributed file inventory
`PACKAGE_MANIFEST.json`	Version, license, framework dependency, and core record hashes
`CITATION.cff`	Dataset authors and version
`LICENSE`	CC BY 4.0 license text
`SHA256SUMS`	Checksums of every distributed file except this checksum file