Introduction
Each observability list has two views: everything, and everything that has been scored. The second is not a filter you configure — it is a sub-tab, and switching to it changes the question the screen is answering from "what happened" to "how good was it".
Where to find it
AI Governance > Agent Registry > open an agent > Observability. Sessions and Traces each carry All and Evaluated; Spans has All only. All is where you land. On a model in the Model Registry, Spans has All and Evaluated.
What "evaluated" means here
A record is evaluated when an evaluator bound to this agent has scored it. That requires three things to have already happened:
An evaluator is bound to the agent — Added on the agent's
Evaluationtab. With none bound, theEvaluatedtab is empty however much traffic there is.The record matched the evaluator's filter — An evaluator can be scoped to a subset. Records outside it are never candidates.
Sampling selected it — Evaluators score a proportion, not everything. An unscored record may simply not have been drawn.
Note: An empty
Evaluatedtab almost never means quality is bad. It means nothing has been scored — no evaluator, a filter that excludes this traffic, or a sampling rate low enough that little has been drawn yet. Check theEvaluationtab's coverage figure before drawing any conclusion from an empty list.
Why the same evaluator does not appear on every list
Evaluators declare what they run on, and the two evaluated lists divide along that line:
List | Shows work from | Because that evaluator |
| Conversational evaluators | Scores the whole conversation |
| Everything that is not conversational | Scores one run at a time |
So an agent with only conversational evaluators bound will show scored sessions and no scored traces — correctly. The two lists are not two views of one dataset; they are populated by different evaluators.
Reading the evaluated sessions list
Column | Reading |
| The scoring outcome for that session |
| The record itself — opens the detail |
| What the session was about |
| How many runs it contained |
| What it consumed |
| How long it took end to end |
| When it began |
The pairing to read across is score against cost. A high-scoring session that cost five times its peers is a quality success and an efficiency problem, and only this list shows both at once.
What this view is good for
Reviewing failures rather than browsing traffic — Sorting the evaluated list by outcome puts the low-scoring runs together. That is a review queue; the
Alllist is not.Sanity-checking a new evaluator — After binding one, come here and read a handful of scored records against your own judgement. An evaluator that disagrees with you consistently needs its criterion rewritten, not more traffic.
Building a test set from real traffic — Scored sessions that went wrong are the best raw material for an offline dataset — they are known-bad cases that actually occurred. See Online vs Offline Evaluation.
"The All tab tells you what your agent did. The Evaluated tab tells you what someone has actually looked at."
The honest caveat
Because evaluation is sampled, the evaluated list is a sample and its scores are a sample's scores. A ninety-percent pass rate over two percent of traffic is a statement about that two percent.
The Evaluation tab reports coverage next to quality for exactly this reason — "{scored} of {total}". Quote the two together or neither. See Sampling & Coverage.