Prospective evaluation and falsifiability protocol

Abstention-aware forecast scoring

Prevent selective silence from inflating apparent accuracy.

Required inputs

All eligible tasks

Prediction or abstention per task

Locked scoring rule

Coverage threshold

Ordered execution

Work the protocol

  1. 1

    Require one submission state for every task.

  2. 2

    Count abstention in coverage.

  3. 3

    Score accuracy both overall and when answered.

  4. 4

    Apply the preregistered abstention penalty.

  5. 5

    Report confidence intervals and missingness.

Outputs

  • Coverage
  • Overall score
  • Conditional score
  • Abstention count

Done only when

  • Tasks equal answered plus abstained.
  • Coverage accompanies accuracy.
  • Policy matches preregistration.

Fail closed

Refusal conditions

  • Abstained tasks vanish.
  • Coverage threshold changes after scoring.
  • Only correct answered tasks are reported.