## Science Subject Evaluation Rules

These rules apply to science questions across any curriculum and question type. They
add subject guidance to the existing metric definitions; they introduce no new metrics
and override nothing in the base or type layers. Apply them to whichever metric each
concern naturally belongs to. Default stance: a scientifically sound item passes — only
flag concrete defects.

### Scientific accuracy (beyond arithmetic) → `factual_accuracy`
- Mechanisms, causation, conservation laws (mass, charge, energy), and directionality
  must be correct, not merely the final number. A computation can be arithmetically
  right yet rest on a wrong scientific premise — that is still a `factual_accuracy`
  defect.
- Quantities must be **physically plausible and dimensionally sane**: units are
  consistent and convert correctly; magnitudes are realistic for the phenomenon; a
  value cannot violate a physical bound (e.g. a negative amount of a conserved
  quantity, a probability/fraction outside its range, an efficiency or yield above
  100%).

### Experimental design & data interpretation → `educational_accuracy`
- When an item involves an experiment, procedure, or data set, the design must support
  the inference it asks for: a controlled comparison varies the intended variable while
  holding others fixed; the data provided must actually be sufficient to reach the
  asked-for conclusion; correlation must not be presented as established causation
  without support.
- A reading-the-data or interpret-the-graph task must give enough information for the
  intended answer to be determinate.

### Misconceptions as the target of difficulty → `reveals_misconceptions`
- Strong science items are built around **known student misconceptions** (e.g.
  conflating mass with weight, heat with temperature, rate with extent, concentration
  with amount). Distractors or expected wrong answers should map to such predictable
  errors rather than to arbitrary or obviously-wrong alternatives.

### Representations → `stimulus_quality` / `factual_accuracy`
- Diagrams, models, and graphs must be scientifically accurate and, where they encode
  quantities, consistent with the item's data (correct scale, labels, conserved counts,
  and species/identities).
