Critical PRs, counted per pull request#
The critical column now counts the share of a model's pull requests with at least one critical finding, shown beside the raw count. The earlier rate divided critical comments by all comments, which made a quiet model look worst: one critical comment across 23 PRs read 100%. It now reads 4%.
Every rate states its sample#
Models that led fewer than five pull requests get a muted chip and a low-sample note instead of a red figure, and models that led none show no rate at all.
Titles and summaries on every PR#
Drill into a model and each pull request shows its title and a one-line summary instead of a bare number. Session stats now break down by model, so the sessions behind a rate sit next to it. Routing cost lives on Model Router.
What we removed#
- The Average Confidence column.
- The Fix Rate by Model chart, in favour of the critical PRs column.
Improvements#
- Model QualityA context-pressure scale on the session table.
- Model QualityComments Fixed is relabelled to say what it counts.
- Model QualityCharts are tidied so each one answers a single question.
Fixes#
- Model QualitySorting by critical rate orders by incidence, not raw comment share.
- Model QualitySession table tail is trimmed so the busiest sessions lead.
- Model QualityA model with one critical PR out of one no longer reads 100% in red.