Faster dashboard, verified memory, and evaluation scores at a glance

New

• Run scores in the suite list. An evaluation run shows its overall score and flags any critical failure without opening it, so a bad run is obvious at a glance. See evaluations.
• Run history on each test case. A row of dots next to a case shows the outcome of its last few runs, oldest to newest, with a relative "last run" time — a case that just started failing, or has always been flaky, stands out without opening it. See evaluations.
• The test case editor warns when a turn still contains placeholder text such as [EMAIL], which an agent cannot act on.

Improved

• The dashboard loads faster. Pages are fetched as you reach them rather than all at once, and a loading screen appears immediately in your chosen theme instead of a blank page. Moving between pages is quicker because the next page is prepared as you hover its link.
• Interrupted page loads recover on their own. A page that fails to load — a flaky connection, or a new version shipping while you have the dashboard open — is retried automatically, and a clear message with a Reload button replaces the blank screen if it still cannot load.
• Analytics is only downloaded once you have made a cookie choice. Nothing is loaded or sent while the choice is outstanding, and declining discards anything recorded up to that point rather than holding it. See security and privacy.
• Remembered facts are checked against the conversation. Each fact an agent stores must quote the customer's own words, and the quote is verified against the transcript before the fact is kept. Anything that cannot be traced back — including a fact drawn from the agent's own reply rather than the customer's — is discarded. Expect fewer remembered facts per conversation, and more reliable ones. See memory.
• Long conversations are read in full when facts are extracted. Previously only the most recent part was used, so something stated at the start could be missed. Extraction of a long conversation uses more of your allowance than a short one.
• Notification templates show encrypted fields masked. A field such as an email address is delivered partly hidden (j•••@a•••.io) instead of being left out of the message entirely, and Settings → Notifications now labels those fields "(masked)" so you know what will arrive.
• Test cases drafted from a real conversation use realistic sample contact details in place of the customer's own, rather than bracketed placeholders that stopped the case from running.
• The agent list loads faster in workspaces with many agents.
• A test run's trace and evaluator detail views now show tool and knowledge source names instead of raw ids, and evaluators that don't already explain themselves in prose — tool correctness, knowledge grounding, latency, reliability, conciseness — get a plain-English line summarising what happened.

Fixed

• Removing someone's platform admin role, or banning an account, now takes effect immediately rather than up to five minutes later.
• Someone added to a workspace could be refused access for several minutes after the invitation was accepted. Access is now granted on their next page load.
• Members of a workspace closed by scheduled deletion could still reach it briefly after it was removed.
• A test case that called a tool in one turn and referred back to the result in a later turn — "are you sure?" — now replays with that memory intact. Previously the later turn ran as if the tool had never been called.