Insights — Analysis Methodology
Wave-0 Freeze
Corpus Profile
| Source | Duolingo | Babbel | Total |
|---|---|---|---|
| 83 | 60 | 143 | |
| YouTube | 113 | 88 | 201 |
| TikTok | 20 | 0 | 20 |
| Total | 216 | 148 | 364 |
Sentiment Scorecard
| Metric | Duolingo | Babbel |
|---|---|---|
| Complaint rate (non-howto base) | 6.0% | 2.8% |
| Praise rate (non-howto base) | 26.5% | 34.0% |
| Cross-brand co-mentions | 11 records discuss both brands | |
Note: No star ratings in corpus — rates are keyword-incidence from organic text, not star-rating averages. Never compare these rates as if they are the same as review scores.
Corroboration — 7/7 findings confirmed
| Finding | Pass 1 (mechanical) | Pass 2 (blind) | Agreement |
|---|---|---|---|
| Streaks = core identity | High-freq unigram | 13 records (6%) | ✓ |
| Free = top love driver | Prominent unigram | 13 records (6%) | ✓ |
| Hearts = top complaint | Cluster detected | 3 records | ✓ |
| Fluency gap = churn | Bigram cluster | 2 records (directional) | ✓ |
| AI controversy signal | Prominent unigram | 8 records (4%) | ✓ |
| Babbel = structured | Top unigram | 13 records (9%) | ✓ |
| Babbel = expensive vs free | Cluster detected | 10 records (7%) | ✓ |
No divergence detected. Confidence: medium — small corpus (364 records per brief spec); directional agreement holds; rates are indicative.
Source Review — 28/28 loci accepted
LLM judge ran autonomously (declared smoke run). All 28 source loci accepted. No false positives detected. 0 rejected, 0 unreviewed. See full table in L3 report Appendix.