open-alternative-jev: both option orders (raw per-decision results, public items only)

author-order-yes-first--* — the ranked v1.2 row. Options listed A. yes, B. no, as the author's yes_no() helper builds them.
reversed-order-no-first--* — our first adapter listed A. no, B. yes. Not ranked: an adapter mistake, not a model weakness. With the options in reverse order the same model scored 21 % instead of 72 % on answer-judging items (the 68 yes/no items of family adequacy from the imported judge set): small models are very sensitive to option order.

Only items whose text is public in datasets/public/ are included here (easy, the 72 original standard/judge items, the 111 public hard items). Held-out and imported items are left out: their per-item predictions would reveal gold answers or redistribute third-party items. Aggregates over all items are in results/v1.2/wip/jevbench-v1.2-wip-results.json.