How to read a version comparison

A comparison view rewards a specific reading order. Headline score last, not first.

Start with new failures

Anything failing in the candidate that passed in the baseline is the reason this review exists. Read those first and in full.

Then fixed failures

Confirm they are fixed for the right reason. A case can pass because the agent now refuses the whole flow, which is not a fix.

Then behaviour changes

Escalation rate, tool usage mix and average turn count. These shift silently and show up in operational costs weeks later.

Headline score last

Use it as a summary you already understand, not as the thing that forms your opinion.

What to take from this

• New regressions before everything else.
• Check that fixes are real fixes.
• Behaviour drift has a cost even when tests pass.