Questions first.
Answers on trial.

Two separate graders score obligation coverage, role and condition fidelity, and evidence discipline. The normalized mean grader score is the answer endpoint. Unsupported claims, false citations, and qualified abstention are recorded separately.

Only supplied definitions may support the answers. A polished answer can still be unhelpful or insufficiently grounded. The raw responses are preserved before grading.

Each reader process handled 36 items sequentially. It never saw the same question twice, but information can carry over between different questions within that process. When both methods supply identical definitions for the current question, an answer difference cannot be attributed to changed current-question definitions; it may reflect reader or scoring variation, or carryover from earlier assigned contexts. Separate contexts also share a model family; this is not an independent human reader study.

The retrieval sign-flip test assumes exchangeable paired signs. Unlike the answer-allocation test, it is not justified by randomized treatment assignment. Neither procedure removes the limits of purposive tasks, shared model errors, or cross-question carryover.

The scorer initially rejected the obligation indices from both graders because they used 1–4 instead of 0–3. After both graders confirmed the order, only the indices were mechanically corrected. Original files were retained; scores and rationales did not change, and no regrading occurred.