comparison / quality

Same questions.
Different models.

Four question lists. Each model answers cold, in a fresh session, and the answers are judged on the same rubric: does it answer or refuse, does it hold one standard when the wording is flipped, and does it get the checkable facts right. Three models are complete.

Models tested

beam Reflection AI · octoface.com/comp/quality/reflection-beam 95% rubric score · 12 refusals in 199 prompts · truth 16/16 · long battery → glm-5.3-flash Z.ai · octoface.com/comp/quality/glm-5.3-flash 93% rubric score · 9 refusals in 193 prompts · truth 16/16 · long battery → muse-spark-open Meta Superintelligence Labs · octoface.com/comp/quality/muse-spark-open 94% rubric score · 4 refusals in 54 prompts · 54/54 prompts · truth 10/10 · short v4 battery →

How the score works

A check is a specific, stated expectation, not a vibe. For paired items the check is whether the model applies the same standard when the wording or the group is flipped; for factual items it is whether the answer contains the fact in the published key. The score is checks passed over checks available.

beam and glm-5.3-flash were scored on the long lists. muse-spark-open was scored on the short v4 battery (same four axes, fewer items). Treat a few points between those two setups as noise. Prompts that never returned are excluded from both sides of the ratio.