Summary

• Adds a new "Understanding Your Baseline Period" section to ml-configuration.md
• Explains how to calculate the comparison baseline period using the formula: baseline_period = number of models per dimension × train every
• Includes example calculations for both default configuration (54 hours) and custom configurations (e.g., 6 hours)
• Clarifies the relationship between individual model training windows and the overall baseline period
• Helps users understand what time period anomaly alerts are compared against

Changes

Added a new section after "Multiple Models and False Positive Reduction" that:

1. Provides the baseline period formula
2. Shows example calculations with default and custom configurations
3. Explains how each model's training window relates to the total baseline
4. Uses Docusaurus admonition syntax for the formula tip

Audit Context

Session Query (Original):
наш текущий конф:
```
[ml]
enabled = yes
dimension anomaly score threshold = 0.5
number of models per dimension = 6
num samples to smooth = 1
minimum num samples to train = 600
train every = 1h
```

когда приходит алерт об аномалии, как понять, по сравнению с чем пришёл этот алерт? за какой период?

Session Query (English translation):
Our current config:
```
[ml]
enabled = yes
dimension anomaly score threshold = 0.5
number of models per dimension = 6
num samples to smooth = 1
minimum num samples to train = 600
train every = 1h
```

When an anomaly alert comes, how to understand what this alert is compared against? For what period?

Decision Verdict: Surgical documentation edit to add a formula and explanation for calculating the ML anomaly detection baseline period based on configuration parameters.

Summary by cubic

Adds an "Understanding Your Baseline Period" section to ml-configuration.md to explain the time window anomaly alerts are compared against. Includes a simple formula and examples to map ML config to a concrete baseline period.

• New Features
  • Defines baseline formula: baseline_period = number of models per dimension × train every.
  • Adds examples: default 18×3h → 54h; custom 6×1h → 6h.
  • Clarifies how per-model training windows relate to the total baseline and that alerts use consensus across all models.

^{Written for commit 72df1a3. Summary will update on new commits.}