Only Tradeoffs: verification bundle for Proposition 42 (Prohibits New State Personal Property Taxes and Certain Retroactive State Taxes), California, November 2026

Everything here is what the models received and returned. To rerun:

  pip install pyyaml jsonschema boto3 google-genai openai
  python3 tools/run_eval.py measure.json --concerns values.yaml --provider <vertex|bedrock|openai> --model <id> --runs 3 --out myruns
  # compare your results (value x result, and clause x value x effect) with reconciled.json

Or paste prompt.txt into any chat and compare the JSON it returns.

Contents
  measure.json          the clauses and current law the models were given
  values.yaml           value list 1-draft.r4
  prompts/              system prompt, user template, blocked terms (v1)
  schema.json           output schema
  runs/                 every attempt by every model: manifest, exact prompt file name, raw reply,
                        parsed output, validation errors (rejected attempts included)
  reconciled.json       the published result
  tools/                the scripts that built the prompts, validated, and reconciled

Models are not fully deterministic and vendors update them, so compare findings, not wording.
