How the dataset is built

This interactive process map follows every number from a frozen source unit to a hashed artifact. Hover over or tap any step to see the exact rule, file boundary, result and validation status.

data artifactindependent model rundeterministic scripthuman judgment

Frozen corpus

113 documents · 3,963 pages

↓

Page unitisation & render routing

hashes · language · text / image mode

↓

OpenAI extraction

5,675 verified statements

Claude extraction

5,683 verified statements

↓

Seal, canonicalise, verify spans

one source-only retry maximum

↓

Deduplicated candidate union

6,949 candidates · 38.0% single-provider

↓

Controlled calibration gate

precision passed · recall failed · spans passed

↓

Equal-vote ensemble

one candidate = one unit of mass

↓

Purpose & privacy measures

6 centres · vocabulary · posture

Statistical analysis

bootstrap · permutation · Holm

↓

Double-blind human audit

independent phase complete · adjudication open

↓

Paper · figures · website · release tables

hash-linked build outputs

Pick a step

Hover over any box on the left. Click to pin the explanation while you inspect its files, rules and audit result.