Results
The Results tab on the agent page shows how the agent is performing across its runs. Open it from the agent page, or with Dashboard in the agent card’s menu.
Choosing what to look at
- Validation runs, Live calls or All calls (the default).
- The period: Last 4 weeks, Last 12 weeks (the default), Last 6 months or Last year.
- Every version, or one workflow version.
- Matrix on <field>: which yes/no output field the confusion matrix uses, when the agent has more than one.
Ground truth comes from labels first, then from verdicts recorded on individual runs.
Tiles
| Tile | What it shows |
|---|---|
| Successful runs | Runs that finished, out of all runs in scope. |
| Labelled | Runs with ground truth or a verdict. |
| Not labelled | Successful runs with no truth. |
| In the matrix | Yes/no answers with a truth. |
| Median run time | Of successful runs. |
Classification
For a yes/no output field, Classification on <field> shows a confusion matrix (Truly yes / Truly no against the agent’s yes and no: true positives, false positives, false negatives and true negatives) and, with their confidence intervals:
| Measure | Meaning |
|---|---|
| Sensitivity | Of the patients who truly are yes, the share the agent said yes to. |
| Specificity | Of the patients who truly are no, the share the agent said no to. |
| Positive predictive value | When the agent says yes, how often it is right. |
| Negative predictive value | When the agent says no, how often it is right. |
| Accuracy | The share of labelled patients the agent got right either way. |
An agent with no yes/no output field has no matrix.
Charts
- Runs per week: live and validation runs each week.
- Accuracy by field: how many labelled answers were right for each output field.
- Recommendation distribution: when the agent has a field named like recommendation, change, action or next step, how often it gave each value.
Ask about these results
The chat on the Results tab answers questions about the agent’s runs. Try the suggestions, such as What do the misses have in common?, Which rule do clinicians overrule most often?, Summarise the edited live calls this month or Are any patients labelled inconsistently?
- Show patients opens the Validation tab on its misses.
- Send to refine flow (editors) opens Refine.