# Cohort analysis

Trace directory: `fixtures/cohorts/before-after`
Group by: `model`
Baseline/Candidate key: `cohort` (before vs after)
Status: **REGRESSION**

## Groups
### after / gpt-4-mini (2 runs)
- Error rate: 50.0%
- Avg duration: 135 ms
- Dominant tools: deleteAccount
- Observation failure rate: 50.0%

### before / gpt-4-mini (2 runs)
- Error rate: 0.0%
- Avg duration: 80 ms
- Dominant tools: searchDocs
- Observation failure rate: 0.0%

## Comparisons
- Error rate: baseline=0 candidate=0.5 (delta 0.5) **REGRESSION**
- Average duration (ms): baseline=80 candidate=135 (delta 55) **REGRESSION**
- Tool choice: baseline=searchDocs candidate=deleteAccount **REGRESSION**
- Observation failure rate: baseline=0 candidate=0.5 (delta 0.5) **REGRESSION**
