docs: add DATA_SPLITS.md (scaled eval sets + experiment→split map)

[overthelex]

Adds DATA_SPLITS.md — the canonical map from experiments to the scaled evaluation sets, so everyone runs against the same splits and fields.

What this documents

• Datasets (public, verdict-free, HUDOC-only, seed=42, leakage-audited v1.1):
  • overthelex/echr-livehrb-static-2k — group ∈ {regular, ukr}
  • overthelex/echr-livehrb-temporal-2k — group ∈ {regular_temporal, ukr_temporal} + bin
• Field reference — verdict_free_text (input), violation_label (gold), article, respondent, group/bin.
• Article buckets — substantive vs procedural (Art. 41 + procedural reported separately, per the MFT-baseline collapse).
• Experiment → split map — baseline, Ukraine effect, temporal drift, UA pre/post 2022-02-24, and the state-swap counterfactual protocol (hold facts fixed, swap respondent state to separate country-prior from temporal drift).

Supersedes the 141-pair pilot for the headline benchmark; the pilot stays valid for the RQ1–RQ3 methodology checks.

cc @TerryJCZhang — flagging for the data-side coordination; happy to adjust bucketing / naming to match repo conventions.

🤖 Generated with Claude Code