SPReAD Project-Type Classification — Reproducibility Archive

Reproducibility archive for the classification of all 1,112 projects funded by Japan's SPReAD program (MEXT AI for Science Emerging-Challenge Research Program, FY2026) into three project types: data-resource construction (D), method/tool development (M), and task response (A). Contents:- Master dataset of the 1,112 funded projects, parsed from the official MEXT award lists (round 1: 22 June 2026; round 2: 18 August 2026), and the PDF parser.- Classification codebook (v0.2) and coding instructions. Project type is judged from the project title only.- Machine classification of all 1,112 projects, with run log.- Human-verification sample of 224 projects (20.1% of the census), with the machine label, two coders' independent labels (machine labels withheld), post-stratification strata, and census weights.- Code that reproduces all agreement statistics (Cohen's κ, unweighted and census-weighted). Key results: on the D/non-D distinction, the two human coders agreed at κ = 0.62, and the machine labels agreed with each coder at κ = 0.52 and 0.73. This archive accompanies the paper "Buying data or buying discovery? What Japan's AI-for-Science grants purchased across 1,112 projects" (preprint: https://doi.org/10.2139/ssrn.7453078). Version 1.1 adds the human-verification data and agreement code. The 221-project verification draw deposited in version 1.0 was never coded; it is kept in the "99_superseded" folder for transparency. See CHANGELOG.md and README.md for details. License: data and documentation under CC-BY-4.0; code (02_code/) under the MIT License.