Reading a scorecard

Open any completed call and you land on its scorecard. Everything on it is derived from the transcript and the channel's rubric, and every finding links back to the transcript turn it came from.

Health score

The number at the top is the health score, from 0 to 100. It starts at 100, loses 15 points per danger fault and 5 per warning fault, and is then scaled by how well the agent captured the required slots and asked the required questions. Info-level faults, such as poor line quality or low transcription confidence, never lower it.

The colour band is the quick read:

Faults

Each fault shows its code, a one-line description, its severity and the moment it happened. Click one to jump to the transcript turn with the evidence highlighted: the two turns where a question was repeated, the gap that counted as dead air, the customer's correction and the agent reply that ignored it.

Faults can be confirmed or marked as a false positive by a reviewer. False positives are excluded from the health score and feed the workspace's fault precision statistics, so the detectors' reliability is visible per code.

The full list of codes, triggers and thresholds is in Fault codes.

Latency

The latency panel plots how long the agent took to start speaking after the customer stopped, for every agent turn. The summary shows the mean, median (P50), P95 and maximum. Turns above the rubric's agent-latency threshold (8 seconds by default) are also reported as AGENT_LATENCY faults; gaps beyond the dead-air threshold (12 seconds) become DEAD_AIR. See Latency.

Slots and questions

The slot table lists every detail the rubric expected the agent to capture: name, phone, budget, appointment window, whatever the template defines. Each slot has a status:

• Captured: the value was obtained.
• Corrected: the caller changed the value at least once during the call; the final value is shown.
• Missing: a required slot that was never captured.
• Not applicable: the slot belongs to a case branch this call did not enter.

A volunteered flag marks slots the caller offered before being asked. The completion rate is the share of required, applicable slots that were captured or corrected.

Below it, the question list shows which required questions the agent actually asked, giving the coverage rate. Both rates feed the health score. See Slot capture and question coverage.

Compliance

If the rubric defines required or forbidden phrases, this section shows each one with a verdict. A required phrase that was never said is a COMPLIANCE_MISS; a forbidden phrase that was said is a FORBIDDEN_PHRASE. Both are danger-level faults.

Sentiment

For English calls, a line chart tracks the caller's sentiment turn by turn on a scale from -1 to +1, with the agent's average alongside. A SENTIMENT_DROP fault marks the first turn where the smoothed customer sentiment crossed into negative territory; if it recovered before the end of the call, the fault is downgraded to info. See Sentiment.

Transcript

The transcript shows each turn with its speaker, timestamps and text, and shades words with low transcription confidence. Faults, slots and sentiment markers are anchored here, and the audio player (when audio is available) seeks to a turn when you click it.

Transcription and line quality

Two cards report how much to trust the transcript itself:

• Confidence aggregates the average word confidence and speaker-attribution confidence. Many low-confidence words trigger an ASR_LOW_CONF info fault; uncertain speaker labels trigger DIARIZATION_UNCERTAIN.
• Line quality is an acoustic score from 0 to 100 (80 and above is clean) built from signal-to-noise ratio, hum, hiss, clipping and dropouts, each of which can raise a LINE_QUALITY info fault. These are properties of the phone line, not of the agent, and they never affect the health score.

See Transcription quality.

Summary and review

A generated call summary sits at the top when summaries are enabled. Marking a call as reviewed removes it from the review queue; Share creates a read-only link to this scorecard for a coach or a client (Shareable scorecards); Mark as golden adds it to the calibration set with the score you believe it deserves.