# Scoring and Grading

The same grader was applied to every arm.

- Answer extraction is identical across arms: the final marked answer is read
  first; a formatting-tolerant rule then reads the last integer so that no arm
  is penalized for style.
- Numeric answers are compared by symbolic equivalence, not string match.
- An arm is scored correct on a task only when its extracted answer is
  equivalent to the frozen gold value.
- Timeouts are recorded, not retried. An arm that exceeds the wall-clock cap on
  a task is marked incomplete for that task and its cost is reported as a lower
  bound (flagged as cost_is_lower_bound in the results file).

Grading never depends on output format, and gold values are never shown to any
arm before or during a run.