The ranking#

At most five causes, rendered {label} — confidence {n}% (confidence stored 0–1, displayed as a percentage), each with an explanation and, when the AI ranking ran, an additional AI-written narrative. Honest fallbacks surface instead of pretending: AI assist ran but produced no narrative. and AI ranking not used: {reason}. The expanded view may also open with a one-line AI summary of the diagnosis.

The evidence#

Items read {kind}: {summary} ({date}) where kind is one of rum_regression_alert, resource_change, configuration_change, origin_change, security_event, cache_purge — the fenced proof tying a cause to what actually changed in the window. When nothing correlates you see No correlated evidence was found in this diagnosis window., and an insufficient evidence status instead of a forced guess.

Reading order that works#

Start from the top cause's confidence, then confirm each one against its evidence items — a high-confidence cause with a configuration_change and a cache_purge on the same day is far more actionable than a bare percentage. The AI narrative (when present) is a summary of that evidence, not a replacement for it: the fallback strings exist precisely so the screen never dresses up a thin result. Expansion state is per row, so you can compare two diagnoses side by side without leaving the table.

Because expansion state is per row and the table never reloads behind your back, you can park two diagnoses open side by side while comparing their cause stacks — useful when a manual run and an alert-driven run cover overlapping windows.

Knowing when a proposal is possible#

Generate proposal only makes sense for completed diagnoses tied to a metric. Without one the expansion explains: No RUM regression metric is attached to this diagnosis … Proposals require a diagnosis triggered by a specific metric (LCP, INP, CLS, or TTFB). Every call is recorded — details in Test a proposal.