Ballot Guide Lab — revised voter questions (v2)
Prepared October 11, 2026. Run complete: 45 outputs, 45 AI ratings, 0 human ratings.

WHAT CHANGED
All 15 main questions now use conversational wording tied to a voter concern.
The set keeps 10 regular and 5 difficult cases. V2-04 replaces the audit keyword
question with a question about waste and accountability. Two abstract questions
are replaced by job-security and everyday-price concerns, grounded in the
Legislative Analyst's page 39 analysis. The former artificial
instruction attack is retained separately as R-01. The original historical
wage-baseline failure is retained separately as R-02. V2-15 instead tests a mixed
question: community-college funding is supported, a local allocation is not.

These are authored examples, not questions collected from users or a measured
ranking of common questions. Ask prospective users how they would phrase their
concerns before claiming this set represents real usage. This test set remains
about the November 5, 2024 election; the existing historical system instructions
remain in force. Candidate quotations are still outside this ballot pilot.

CASE FORMAT
The primary JSON is a 15-element array using the original runner fields: id,
proposition, difficult, question, expected_status, reference_pages, answer_notes.
Extra intent and wording_rationale fields help human reviewers. V2 case IDs avoid
confusion with the archived v1 cases. R-01 and R-02 belong to a separate regression suite.

NEW PROJECT ZIP
prototype-v2-voter-questions.zip contains the original portable code with only:
- pilot/cases.json replaced by this v2 set;
- pilot/cases_v1.json retained as an archival reference;
- pilot/regression_cases.json holding the instruction-attack and wage-baseline regressions;
- Colab pipeline/notebook output directories changed to colab_lora_v2_voterquestions;
- old generated result directories omitted from this new runnable package;
- notebook outputs and execution counters cleared.
The base checkpoint, model prompts, training data, retrieval settings, rubric,
training recipe and decoding settings are unchanged. No adapter weights are
included. The adapter remains on the pinned Hugging Face repository.

REPRODUCING THE RUN
Use a new directory and fresh outputs. The included Colab notebook follows the
original training recipe; it does not activate any hosted paid endpoint. Fresh
GPU execution is still needed. If reusing the already trained adapter instead
of retraining, record its pinned revision and exact loader/decoding settings.
Run all 15 cases in the naive untuned, source-only untuned and source-only
fine-tuned conditions, with identical per-question evidence across conditions.
Changed wording can change BM25 retrieval; inspect the retrieved passages too.
Freeze and blind the 45 new outputs before independent AI and human ratings.
Apply the original four-criterion rubric and report actual disagreements.
Do not copy the old 2/15, 3/15 or 7/15 results onto the revised questions.

Within the extracted project, the scoring preparation command is:
python evaluate.py prepare --run-dir results/colab_lora_v2_voterquestions

The ordinary local run_pilot.py supports the same v2 case file, but runs only
two prompt conditions on its configured local backend. Its outputs must not be
reported as the trained 2B three-condition Colab experiment.

FILES
The original v1 archives, report and human review workbook remain unchanged.
Use new blinded outputs and new human-rating templates for v2. Do not overwrite
or reinterpret already frozen answers. Current v2 outputs: 45; independent AI scores: 45; human scores: 0.
Use Ballot_Guide_V2_Run.ipynb to reuse the pinned saved adapter without training.
Read v2-review-instructions.txt for the human-rating export commands.
The two optional regression cases were not run.