LegalBench Abercrombie dataset-statistics check — 2 October 2026

Scope: dataset composition only. No model performance or legal labels were tested.
The report was captured before publication; its assurance fields describe that
local execution stage. The public export removes workstation paths from the log
and execution receipt. Scientific values and input hashes are unchanged.

Expected test-class counts in arXiv:2308.11462v1 Table 7: 19,19,20,18,19.
Pinned artifact: 19 in each of five classes; test total 95, training total 5.
The historical experiment's exact data revision is not established. This mismatch
does not establish an error in reported model performance or its cause.

Attribution: Neel Guha and LegalBench contributors, Guha et al. (2023).
Task dataset licence: CC BY 4.0, https://creativecommons.org/licenses/by/4.0/
Source paper: https://arxiv.org/html/2308.11462v1#S5.T7
No original paper or third-party dataset is redistributed in this download.
The scripts were written by the research assistant; upstream code is not executed.

To repeat on macOS/Linux with Python 3, in a fresh copy of this directory:
  mkdir inputs logs
  mv report.json original-report.json
  mv input-manifest.json original-input-manifest.json
  python3 -I fetch_inputs.py > logs/download.stdout.jsonl 2> logs/download.stderr.log
  python3 -I run_audit.py report.json

Compare fetched hashes against original-input-manifest.json before interpretation.
Code/data commits are pinned; hosting metadata and regenerated HTML may change.
The initial and replay measurements used identical locally retained input bytes.
No pip packages or paid model calls are needed. CPU and wall limits are applied;
the requested memory limit was not enforceable on the original macOS host.
The scripts execute locally, not inside a security sandbox.
