Sync SKILL.md: start anomaly detection from charm_knn_scores / charm_zero_shot_scores
#3
by gerar444 - opened
- skills/charm/SKILL.md +3 -2
skills/charm/SKILL.md
CHANGED
|
@@ -71,8 +71,9 @@ These patterns are starting points; combine them:
|
|
| 71 |
- **Detect without labels.** Score windows by distance to normal behavior. Add cheap window
|
| 72 |
statistics, because normalized embeddings miss level shifts and spikes. With no clean
|
| 73 |
reference, bootstrap one — and show the user *which* periods were selected as normal before
|
| 74 |
-
trusting the scores.
|
| 75 |
-
|
|
|
|
| 76 |
- **Embeddings as features.** Feed vectors, plus window statistics and metadata you already have,
|
| 77 |
into sklearn models. This is the fastest supervised baseline.
|
| 78 |
- **Escalate to a trained head only when needed.** If a zero-shot or sklearn baseline falls short
|
|
|
|
| 71 |
- **Detect without labels.** Score windows by distance to normal behavior. Add cheap window
|
| 72 |
statistics, because normalized embeddings miss level shifts and spikes. With no clean
|
| 73 |
reference, bootstrap one — and show the user *which* periods were selected as normal before
|
| 74 |
+
trusting the scores. Start with `charm_knn_scores` (clean reference) or
|
| 75 |
+
`charm_zero_shot_scores` (none): each runs the best recipe in one call. `anomaly-detection.md` explains
|
| 76 |
+
them and the lower-level building blocks, with a symptom→knob tuning guide and an evaluation snippet.
|
| 77 |
- **Embeddings as features.** Feed vectors, plus window statistics and metadata you already have,
|
| 78 |
into sklearn models. This is the fastest supervised baseline.
|
| 79 |
- **Escalate to a trained head only when needed.** If a zero-shot or sklearn baseline falls short
|