WEBVTT

00:00:00.000 --> 00:00:07.500
Two configurations. All 48 original answers.
Qwen3-8B BF16 and Qwen3.8-27B FP8; three fresh answers per public development case.

00:00:07.500 --> 00:00:15.000
The total improves. The upgrade is blocked.
Complete plans: 15/24 → 19/24. Ten named checks still lose passes.

00:00:15.000 --> 00:00:22.500
Inspect a regression, not just the total.
already-closed · seed 17: the current answer omits the required actions key.

00:00:22.500 --> 00:00:30.000
Reproduce the decision from native results.
Compare the downloaded JUnit files with evalarc==0.15.0. Expected gate exit: 1.