Criteria Writing Guide
Reward / Penalize / Passes if / Fails if — reliable criteria patterns.
Criteria Writing Guide
Criteria are the instructions QAVE scores against. Prefer one focused quality with clear direction.
Phrasing styles
| Style | Phrasing | Use when |
|---|---|---|
| Continuous | Reward … / Penalize … | Graded 0–1 along one quality |
| Binary | Passes if … / Fails if … | Clear pass/fail threshold |
Principles
- One quality per criterion — split unrelated checks.
- Clear direction — start with Reward / Penalize / Passes if / Fails if.
- Monotonic or qualified — use an
if …qualifier to scope applicability. - Avoid conjunctions —
and/orusually smuggle a second quality. - Avoid role-keyword collisions — do not embed
user:/assistant:role tokens inside criteria text.
Template
[Prefix] [quality] [qualifier]
Example: Reward responses that cite the exact policy section if the user asks about cancellations.
Good examples
Reward responses that state the refund window in days.Penalize responses that invent policy details not present in the input.Passes if the response refuses to give medical dosage advice.
See also the Criteria Library for packaged presets.