Platform/Data Quality
Data-Quality Controls
Every metric has definition, formula, owner and freshness. Missing data is never silently treated as zero.
Principle: never silently treat missing data as zero. Show quality scores on integrations and calls.
Issue classes open
10
Catalog metrics
4
Integrations scored
5
Detected issues (24h)
| Type | Count | Severity |
|---|---|---|
| Duplicate calls | 3 | medium |
| Missing timestamps | 12 | high |
| Invalid event order | 2 | medium |
| Incomplete transcripts | 28 | high |
| Recording mismatch | 1 | critical |
| Clock skew | 7 | medium |
| Missing provider IDs | 4 | high |
| Impossible durations | 0 | low |
| Outlier values | 19 | low |
| Schema changes | 1 | medium |
| Partial traces | 41 | medium |
Integration data-quality score
Retell AI96 · 2 issues
Vapi94 · 3 issues
Twilio91 · 5 issues
LiveKit93 · 4 issues
Generic OTel88 · 9 issues
Metric catalog
Definition · formula · unit · source · owner · version · freshness · limitations
| Metric | Formula | Unit | Owner | Ver | Freshness | Limitations |
|---|---|---|---|---|---|---|
Task success rate Share of completed calls with successful business outcome | successful_calls / completed_calls | % | Quality | 1.3 | 1m | Depends on outcome mapping; missing outcomes excluded (not zero) |
P95 end-to-end latency 95th percentile user-perceived turn latency | percentile(turn.e2e_ms, 0.95) | ms | Platform | 2.0 | 30s | Clock skew can inflate values |
Containment rate Calls resolved without human transfer | 1 - transferred_calls / completed_calls | % | Ops | 1.1 | 1m | Warm transfers may be misclassified if provider omits events |
Tool error rate Tool spans ending in error or timeout | tool_error_spans / tool_spans | % | Engineering | 1.0 | 30s | Partial traces under-count tools |