VetoBench / Changelog

Every material rule change gets a date.

The draft will change. This page records changes that affect prompt construction, scoring, access, adjudication or the meaning of a published metric.

Public project draft

  • Defined the initial safe, control and extension sets.
  • Separated provider errors, explicit refusals, generic refusals and task misses.
  • Set three repeated runs per case for the first pilot draft.
  • Published the provider-access and independence position.