cekura-predefined-metrics

Original：🇺🇸 English

Translated

Use when the user asks "what predefined metrics are available", "which built-in metrics should I use", "what does CSAT measure", "how does hallucination detection work", "what's the difference between Interruption Score and AI Interrupting User", "which metrics are free", "which metrics need audio", "configure silence threshold", "set up sentiment metric", or any question about Cekura's out-of-the-box metrics. Covers the full catalog of predefined metrics — what each does, costs, constraints, configuration options, and when to use each one.

2installs

Sourcecekura-ai/cekura-skills

Added on2026-05-31

NPX Install

npx skill4agent add cekura-ai/cekura-skills cekura-predefined-metrics

SKILL.md Content

View Translation Comparison →

Cekura Predefined Metrics

Purpose

Predefined metrics are Cekura's built-in evaluators — ready to enable on any agent with no prompt writing required. They cover the most common quality dimensions across accuracy, conversation quality, customer experience, and speech quality. Use this skill to decide which predefined metrics to enable and how to configure them.

Performing Platform Actions

When this skill suggests creating, listing, updating, or evaluating something on Cekura, prefer using available platform tools over describing API calls or dashboard steps. In Claude Code with the Cekura plugin installed, these tools are auto-configured and handle authentication, parameter validation, and error handling for you. Fall back to direct API endpoints or dashboard guidance only when no tools are available in the current session.

Core Terminology

Main agent: The client's AI voice agent being tested
Testing agent: Cekura's simulated caller that exercises the main agent
Predefined metric: Built-in evaluator shipped by Cekura — no prompt required, identified by a
```
code
```
Custom metric: User-authored metric with a custom prompt or
```
custom_code
```
(see
```
cekura-metric-design
```
)
Simulation: Test runs using Cekura's testing agent against the main agent (Sim column in the catalog)
Observability: Real production calls flowing through the agent (Obs column in the catalog)
Project-level toggle: Enables a predefined metric across simulation OR observability for an entire project
Evaluator attachment: Adds the metric to a specific test scenario; required for the metric to fire on that evaluator

Predefined vs Custom Metrics

Enable predefined metrics first. They require zero prompt engineering and cover the most common quality dimensions out of the box. Only reach for a custom metric when:

You need to evaluate a business-specific workflow (booking flow, escalation protocol, etc.)
You need to check agent behavior against your specific system prompt
You need to combine multiple signals into one score

For everything else, a predefined metric will give you reliable, consistent results faster.

The Predefined Metrics Workflow

Browse the catalog — Use the four catalog tables below to identify candidate metrics. Filter by Sim/Obs availability, cost, and required constraints.
Pick a starting set — Begin with the Baseline below. For richer coverage, see
```
references/selection-by-use-case.md
```
for recommended sets per agent type (booking, collections, support, healthcare, voice-quality investigation).
Toggle on at the project level — Enables the metric for simulation runs (or observability — they are independent toggles). Without this, attaching the metric to an evaluator does nothing.
Add to individual evaluators — Attaches the metric to specific test scenarios so it fires when that scenario runs.
Configure if required — Six metrics need or accept configuration (silence thresholds, dropoff/topic node lists, spelling categories, pronunciation phonemes). See
```
references/configuration-guide.md
```
for full payload examples.
Validate by running — Execute a small batch and review results. If results look off, check the Common Pitfalls below before reaching for custom metrics.

For full API endpoints (list, toggle, attach, configure, re-evaluate), see

references/api-reference.md

Enabling Predefined Metrics

Two steps are required — missing either means the metric never fires:

Toggle on at the project level — enables the metric for simulation runs
Add to individual evaluators — attaches the metric to specific test scenarios

Use

GET /test_framework/v1/predefined-metrics/

to retrieve the full list of available predefined metrics and their IDs. Pass a predefined metric's

code

field when adding it to an agent's metric set.

Catalog: Accuracy

Metric	Output	Cost	Sim	Obs	Notes
Expected Outcome	0–100 score	Free	✓	—	Requires `expected_outcome_prompt` set on the evaluator. Scores how well the agent achieved the scenario goal. Without this, runs only pass/fail on call completion. Transcript-only — cannot evaluate voice characteristics (tone, pronunciation, speech quality). When writing the prompt, refer to speakers as "main agent" and "testing agent" — not "user", "assistant", "bot", or "AI".
Hallucination	True/False	0.6 credits	✓	✓	Compares agent responses against the Knowledge Base to detect unsupported claims.
Mock Tool Call Accuracy	0–100 score	Free	✓	—	Scores whether the right mock tools were called with the right inputs. Requires mock tools configured on the agent.
Relevancy	True/False	0.2 credits	✓	✓	Checks if agent responses addressed the question asked. Flags off-topic or deflecting replies.
Response Consistency	True/False	0.2 credits	✓	✓	Detects contradictions — when the agent repeats information incorrectly or contradicts a prior statement.
Tool Call Success	True/False	Free	✓	✓	Checks if any tool call result contains "Error" or "failed". Requires provider integration (assistant ID + API keys) so tool call data appears in the transcript.
Transcription Accuracy	0–100 score	Free for simulations / 1 credit/min for production call logs	✓	—	Uses two transcription models for production call logs, compares against ground truth for runs. Requires audio. Production call log evaluation is expensive — use selectively.
Voicemail Detection	True/False	0.2 credits	✓	✓	Detects if the call reached a voicemail or automated system. Beta.

Catalog: Conversation Quality

Metric	Output	Cost	Sim	Obs	Notes
AI Interrupting User	Count	Free	✓	✓	Counts how often the agent interrupted the user. For observability, requires stereo audio with separate speaker channels.
Appropriate Call Termination by Main Agent	True/False	0.2 credits	✓	✓	Checks whether the agent ended the call prematurely and whether the user's concern was resolved.
Appropriate Call Termination by Testing Agent	True/False	0.2 credits	✓	✓	Checks if the user (testing agent) ended the call abruptly — a signal of poor experience or unresolved issues.
Detect Silence in Conversation	True/False	Free	✓	✓	Returns False if neither speaker speaks for longer than `silence_duration` seconds. Default: 10s. Configurable.
Infrastructure Issues	True/False	Free	✓	✓	Returns False when the main agent goes silent for longer than `infra_issues_timeout` seconds. Default: 10s. Configurable. Distinct from Detect Silence — this is agent-specific.
Interruption Score	0–100 score	Free	✓	✓	Continuous score for how often the agent interrupts the user. Higher = fewer interruptions = better.
Latency (in ms)	ms average	Free	✓	✓	Average response latency. Also reports P25/P50/P75/P90/P95/P99 percentiles. Under 2000ms is considered good.
Stop Time after User Interruption (ms)	ms	Free	✓	✓	Time from user interruption until the agent stops speaking. Lower = more responsive.
Unnecessary Repetition Count	Count	0.2 credits	✓	✓	Counts how many times the agent unnecessarily repeated itself.
Unnecessary Repetition Score	0–100 score	Free	✓	✓	Continuous score for repetition quality. Higher = more concise = better. Prefer this over the count metric for trend tracking.
User Interrupting AI	Count	Free	✓	✓	Counts customer interruptions of the agent. High counts signal frustration or poor turn-taking.

Catalog: Customer Experience

Metric	Output	Cost	Sim	Obs	Notes
CSAT	0–100 score	0.2 credits	✓	✓	Overall customer satisfaction. Scores above 70 indicate satisfaction. Evaluates tone, cooperation, and resolution.
Dropoff Node	Enum	0.2 credits	—	✓	Identifies the conversation stage where the call ended. Requires `dropoff_nodes` configuration with predefined stage names. Observability only.
Sentiment	Enum	0.2 credits	✓	✓	Classifies user sentiment as Happy, Angry, Neutral, or Disappointed based on tone and word choice across the call.
Topic of Call	Enum	0.2 credits	—	✓	Categorizes what the call was about (e.g., billing, technical support). Requires `topic_nodes` configuration. Observability only.

Catalog: Speech Quality

Metric	Output	Cost	Sim	Obs	Notes
Average Pitch (in Hz)	Hz	Free	✓	✓	Average vocal pitch of the main agent during the call. Useful for monitoring voice consistency.
Gibberish Detection	True/False	0.3 credits/min	✓	✓	Detects garbled or incoherent speech. Requires stereo audio. Beta.
Letterwise Pronunciation Detection	True/False	0.2 credits	✓	✓	Checks if the agent spells things out letter-by-letter when appropriate (e.g., confirming phone numbers). Requires `spelling_word_types` configuration.
Pronunciation Check	0–100 score	0.2 credits	✓	✓	Custom word accuracy — compares spoken output against a list of expected phonemes. Requires `pronunciation_words` configuration as phoneme pairs. Beta.
Speaking Rate	True/False	0.2 credits	✓	✓	Detects abrupt changes in the agent's speaking pace. English only. Beta.
Talk Ratio	0.0–1.0	Free	✓	✓	Ratio of agent speaking time vs user speaking time. Typical healthy range: 0.4–0.6. Requires stereo audio for observability.
Voice Change Detection	True/False	0.2 credits	✓	✓	Detects if the agent's voice changes unexpectedly (different speaker, voice model issue). Beta.
Voice Tone + Clarity	0–100 score	0.2 credits	✓	✓	Audio quality score — analyzes clarity and jitter. Scores above 70 indicate quality.
Words Per Minute (WPM)	WPM	Free	✓	✓	Speaking speed of the main agent. Useful baseline alongside Average Pitch and Talk Ratio.

Configuration Reference

Some predefined metrics require or support configuration. Pass these as key-value pairs in the metric's

configuration

object.

Metric	Config Key	Type	Default	Description
Detect Silence in Conversation	`silence_duration`	int (seconds)	10	Silence threshold for either speaker
Infrastructure Issues	`infra_issues_timeout`	int (seconds)	10	Silence threshold for the main agent only
Dropoff Node	`dropoff_nodes`	array of strings	required	Conversation stage names (e.g., `["greeting", "verification", "booking", "closing"]` )
Topic of Call	`topic_nodes`	array of strings	required	Topic categories (e.g., `["billing", "technical_support", "cancellation"]` )
Letterwise Pronunciation	`spelling_word_types`	array of strings	required	Word categories to check (e.g., `["phone_number", "confirmation_code"]` )
Pronunciation Check	`pronunciation_words`	array of objects	required	Phoneme pairs: `[{"word": "Cekura", "phoneme": "sɛˈkjʊrə"}]`

For full payload examples (including IPA tips and naming guidance) see

references/configuration-guide.md

Cost & Credits Quick Reference

Cost	Metrics
Free (0 credits)	Expected Outcome, Tool Call Success, Mock Tool Call Accuracy, AI Interrupting User, User Interrupting AI, Stop Time after User Interruption, Latency, Detect Silence, Infrastructure Issues, Interruption Score, Unnecessary Repetition Score, Average Pitch, Talk Ratio, Words Per Minute
0.2 credits/call	Relevancy, Response Consistency, Voicemail Detection, Appropriate Call Termination (both), Unnecessary Repetition Count, CSAT, Dropoff Node, Sentiment, Topic of Call, Letterwise Pronunciation, Pronunciation Check, Speaking Rate, Voice Change Detection, Voice Tone + Clarity
0.6 credits/call	Hallucination
0.3 credits/min	Gibberish Detection
Free (simulations)	Transcription Accuracy (simulation runs only)
1 credit/min (production call logs)	Transcription Accuracy (production call log evaluation)

Baseline — Always Enable

At minimum, every agent should have these four enabled for simulation (and the last three also for observability):

Metric	Why
Expected Outcome	Without this, runs only tell you if the call completed — not if the agent actually did the right thing
Infrastructure Issues	Catches the agent going silent for 10+ seconds — invisible in pass/fail
Tool Call Success	Detects broken integrations before they impact real users
Latency	Baseline performance tracking; P95/P99 reveal outliers that averages hide

For a richer baseline, also add: CSAT, Sentiment, and Unnecessary Repetition Score.

For agent-type-specific recommended sets (booking, collections, support, healthcare, voice-quality investigation), see

references/selection-by-use-case.md

Key Constraints

Audio required: Transcription Accuracy and Gibberish Detection need audio data. Not available for text-only runs.
Stereo required (observability): AI Interrupting User, User Interrupting AI, Talk Ratio, and Gibberish Detection require stereo recordings with separate speaker channels for observability calls.
Simulation only: Transcription Accuracy, Mock Tool Call Accuracy, Expected Outcome.
Observability only: Dropoff Node, Topic of Call.
English only: Speaking Rate.
Requires configuration: Dropoff Node, Topic of Call, Letterwise Pronunciation Detection, Pronunciation Check — will not produce meaningful results without the configuration keys set.
Requires provider integration: Tool Call Success requires the agent's provider assistant ID configured on Cekura so tool call data appears in transcripts.

Common Pitfalls

Enabling metrics without completing both activation steps (project toggle AND evaluator assignment) — metrics appear available but never fire
Using Detect Silence and Infrastructure Issues interchangeably — they measure different things (both speakers vs agent only)
Expecting Transcription Accuracy on observability calls — it's simulation-only
Forgetting
```
expected_outcome_prompt
```
when using Expected Outcome — without it the metric has nothing to evaluate against
Using Expected Outcome to evaluate voice characteristics (tone, pronunciation, speech quality) — it only has access to the transcript; use Speech Quality metrics for audio evaluation
Writing
```
expected_outcome_prompt
```
with terms like "user", "assistant", "bot", or "AI" — always use "main agent" and "testing agent" to match Cekura's transcript labeling
Enabling Dropoff Node or Topic of Call without configuring
```
dropoff_nodes
```
/
```
topic_nodes
```
— results will be meaningless
Using Gibberish Detection on mono recordings — it requires stereo audio

Next Steps

After selecting predefined metrics, the user typically needs:

Create or configure metrics → invoke cekura-metric-design for custom metrics to complement predefined ones
Improve a metric that's underperforming → invoke cekura-metric-improvement for the feedback and labs cycle
Attach metrics to test scenarios → invoke cekura-eval-design to wire up metrics in evaluators

Documentation

Pre-defined metrics reference: https://docs.cekura.ai/documentation/key-concepts/metrics/pre-defined-metrics
Public docs: https://docs.cekura.ai
Concepts: https://docs.cekura.ai/documentation/key-concepts/

Additional Resources

Reference Files (loaded on demand)

references/configuration-guide.md
— Full payload examples for every configurable predefined metric (silence thresholds, dropoff/topic node lists, spelling categories, IPA phoneme pairs)
references/api-reference.md
— Public API endpoints for listing, toggling, attaching, configuring, and re-evaluating predefined metrics, with an end-to-end example flow
references/selection-by-use-case.md
— Recommended predefined metric sets by agent type: booking, collections, customer support, healthcare, and voice-quality investigation