Analysed call
A call SignalMD has finished reading, so it has an outcome and, where something went wrong, what went wrong. Board numbers and version comparisons count analysed calls. Your plan’s credits pay for processed minutes: each minute of a call uses one credit once it has started. See How SignalMD works.Brief
The short document that says what your agent is for and what success means. SignalMD drafts it from your agent’s configuration, and you edit and approve it. It shapes how issues are ranked and what counts as designed behaviour. See The brief.Call
One conversation between a caller and your agent, imported from Retell with its full record. Open one under Calls to see its recording, transcript and what went wrong. See Calls and flow.Caller
The person on the other end of a call. SignalMD reads who they are and why they called. A caller who never speaks, or who hangs up in the opening seconds, is not an issue: no change to your agent would stop it, so it counts in the call outcomes and is in neither Fix nor Improve. See Lanes.Cause not established
SignalMD’s answer when a problem is real but the calls do not show why. SignalMD does not guess: these problems are listed on Issues under their own heading, never under Fix, and the issue page says what to check instead of proposing a change. See Cause not established.Comparison
A before-and-after reading of production calls, between two versions of your agent or between two date ranges, with each metric checked for a real change. It is not a controlled experiment, because Retell ships a version to all traffic at once. See Versions.Coverage
How well your tests cover what callers ask for. The Coverage tab of Tests marks each use case, kind of caller and open issue Covered, Partly or Not covered. See Test a fix.Evidence
On an issue, the distinct quotes from its calls, each linking to that moment in the call. On a call, Evidence for lists the issues the call belongs to. See Issues.Fix
The board section for confirmed defects: problems SignalMD has confirmed across calls and whose cause it established, in your agent or outside it. When the cause is outside your agent, the row says “Outside your agent”. Overview shows the top 3. Problems whose cause is not established are listed apart, under Cause not established on Issues. See Fix and Improve.Flow
The steps your agent moves through and the transitions between them: the nodes of a conversation flow, or the states of a single-prompt agent. The Flow page draws it with real traffic on it, one agent version at a time. See Calls and flow.Flow and state
A lane. What your configured flow did with what was understood: asked again, skipped a question, looped, stored an answer against the wrong question, or kept an old value. Usually fixed in the flow’s design. See Lanes.Got it done
The share of calls where the caller got what they came for, as SignalMD’s reading of each call judged it. It is the primary measure on the overview and in version comparisons, and it leaves out calls where the caller never spoke. See Versions.Grace period
How long your agent must take calls without an issue before SignalMD counts the issue as resolved, and how long a new version must run before SignalMD says whether it fixed an issue. 7 days by default, set per agent on Issues to 1, 3, 7, 14 or 30 days. See The grace period.Hearing
A lane. What the caller said compared with what was recorded: a misrecognised value, a dropped “not”, a caller cut off mid-sentence. Usually fixed in speech recognition settings and vocabulary. SignalMD does not measure recognition accuracy directly yet. See Lanes.Improve
The board section for changes you control, in a prompt, the flow, routing or a tool setting, that move one of your numbers without being one malfunction. Each row says the change and the number it moves. See Fix and Improve.Integrations
A lane. Tools and knowledge bases that fail, hang, return nothing, or return something the flow cannot use. Usually fixed in the tool or the system behind it. See Lanes.Issue
One problem you can fix. It names a place, a cause and a change to make, and it carries the calls that show it. Calls are grouped into one issue when one change fixes them all. See Issues.Lane
One of five places a defect can live: Hearing, Understanding, Flow and state, Speaking and Integrations. The lane tells you who owns the fix. See Lanes.Not comparable
A verdict meaning the two sides were read by different versions of SignalMD’s analysis, so part of any difference could come from SignalMD rather than your agent. It clears once the older calls are read again. In a run, the same label marks a test whose result cannot be compared with the earlier run. See Versions.Persona
A kind of caller read from real calls and described by five traits: Who they are, How they behave, How they sound, Around them and Language. A test can play a persona, so a fix is checked against the callers you actually get. See Test a fix.Place
Where an issue happens: a step, a tool or a transition. Every issue names one. The issue page shows it under the title, or after Where under the cause. See Issues.Priority
How SignalMD orders issues: by impact, adjusted by what your brief marks as mattering most. Treat the order as a starting point and read the calls. See Issues.Release set
The tests SignalMD runs on every new version, and the ones a CI check runs. A test is in it when marked In release on Tests; replays of real calls join it when you accept them. You can also give tests a group name and run a group together. The API lists these as test sets. See Test a fix.Run
One play of tests against one version of your agent, as simulated calls. Each test comes back Passed, Failed, Flaky, Error or Caller drift. Runs are listed under Runs, and the API calls each simulated call in a run an attempt. See Test a fix.Signal (check)
A small observation about one moment in a call, such as an agent saying a vehicle was added when nothing in the call added it. Signals appear on the call page. A signal is evidence for a diagnosis and never an issue on its own. See Signals.Speaking
A lane. What the caller heard and when they heard it: a long wait before a reply, the agent talking over the caller, turns that run long. Usually fixed in the model choice, prompt length, or voice and responsiveness settings. See Lanes.Step
One node of a conversation flow, or one state of a single-prompt agent. A shared component counts as one set of steps wherever you placed it. See Calls and flow.Test
A simulated call you play against a version of your agent before it reaches callers. It usually comes from a real call behind an issue, with the original caller’s goal and behaviour, the tool responses that call got (personal details masked) and criteria that include the issue not happening again. Tests run as text conversations. See Test a fix.Too small to matter
A verdict on one metric: the movement is real, but too small to act on, so SignalMD treats it as unchanged. See Versions.Trace
SignalMD’s record of one call as you see it on the call page: what was said, in which step of your flow, what each tool did, and how long each part took. See Calls and flow.Understanding
A lane. What the model did with what it received: instructions not followed, a question misread, an answer invented, content the knowledge base does not support. Usually fixed in the step’s prompt, the global prompt or the knowledge base. See Lanes.Verdict
The label SignalMD puts on a comparison. On one metric it is a signed difference such as +10.0 points, Too small to matter, Within noise, Not enough calls or Not comparable. On a version change it is Observed improvement, Observed regression, Mixed result, No better or worse, Not enough calls or Not comparable. See Versions.Version
One published state of your agent in Retell. SignalMD captures each new version’s configuration when it sees it and compares production calls before and after. A version whose configuration Retell never gave back, such as one that ran before you connected, is marked “no config” in History, and its comparisons are tagged Calls only. See Versions.Next steps
How SignalMD works
From your first calls to a fix that worked.
Lanes
The five places a defect can live.