Chapter 2 · 2026-10-01 → 10-03
Building a ruler
Every answer is graded twice, blind, against a frozen frontier answer, under rules written before the test.
In this chapter · 8
- frontier reference answerconcept
- pairwise judge (both orders)concept
- frozen criterion (pre-registration)concept
- exposure ledgerconcept
- held-out (blind) setconcept
- V-paraconcept
- Freeze every deciding criterion (sha256)…decision
- Every touch of a held-out set was loggedfinding