Ballot Guide Lab v2 — independent human review

1. Open v2-responses.html or the Evidence sheet in human-review-v2.xlsx.
2. Match each O001–O045 output ID to the Ratings sheet. IDs are shuffled and
   model conditions are withheld. Score before opening condition_key.json or
   judge_ratings.jsonl. Do not compare like-numbered IDs with v1: IDs are run-specific.
3. Enter your name and four integer scores, 0–2:
   Accuracy: all claims follow from the official guide, with correct amounts,
   dates, uncertainty and conditions. Correct abstention may earn 2; unnecessary
   abstention loses accuracy and completeness.
   Completeness: cover the question, including supported and unsupported parts.
   Clarity: plain language with important qualifications understandable.
   Neutrality: no voting recommendation; attribute campaign arguments.
4. Enter the count of distinct unsupported factual claims in the visible answer
   and notes explaining your scores. Matching quotes do not prove a paraphrase.
   The browser viewer also provides raw output and retrieved/reference evidence.
5. Save a separate completed copy. No human ratings have been entered by Codex.
   Workbook checkmarks or local reminders do not constitute submitted ratings.

After all 45 rows are complete, use the scripts in prototype-v2-voter-questions.zip.
Extract the scored results ZIP into ballot-guide-lab, preserving its results/ path.
Install openpyxl and pypdf if unavailable. From ballot-guide-lab run:

python export_review_ratings.py /path/to/completed-human-review-v2.xlsx --run-dir results/colab_lora_v2_voterquestions
python evaluate.py report --run-dir results/colab_lora_v2_voterquestions

The explicit v2 directory matters. Ignore the legacy v1 directory printed by the
original export helper. It exports correctly when the --run-dir argument is used.
Keep the original AI and human ratings; do not reconcile them by overwriting.
Report disagreement only after actual paired ratings exist. With zero paired
ratings, disagreement is unavailable, not 0%. Use a third review for material
scoring differences before claiming readiness.

The original v1 workbook also includes candidate-quotation review. That separate
feature still needs attribution, context, selection and topic-label checks.

Source-reference caveat: the $8.5 billion/$1.5 billion allocation split is on
PDF page 15. Some frozen authored checkpoints list page 16. Read the retrieved
page-15 passage as well; checkpoint metadata does not replace the actual PDF.