main
md 34 lines 1.54 KB
Rendered Raw
1 # Skill: Noise Classification
2
3 ## Purpose
4 Identify and filter coordinated manipulation patterns in GitHub crawl data before
5 editorial analysis, preventing inflated repos from contaminating trend signals.
6
7 ## Detection Heuristics
8
9 ### Star-Farming (W22 pattern)
10 - **Trigger**: ≥3 repos landing within ±10 stars of each other in the same crawl window
11 - **Confirm**: Zero or near-zero fork count on repos with 400+ stars
12 - **Context**: Often clustered in game-cheat, software-unlock, or AI-branded categories
13 - **Action**: Flag as `noise:star-farm`, exclude from trend aggregation
14
15 ### Fork-Inflation (W23 pattern)
16 - **Trigger**: Fork/star ratio > 10:1 on repos < 30 days old
17 - **Confirm**: Keyword-stuffed descriptions, bot-pattern naming conventions
18 - **Context**: Polymarket, crypto, and automated-trading verticals
19 - **Action**: Flag as `noise:fork-inflation`, exclude from trend aggregation
20
21 ### Creation-Timestamp Clustering
22 - **Trigger**: ≥5 thematically similar repos created within a 60-minute window
23 - **Confirm**: Similar description templates, near-identical READMEs
24 - **Action**: Flag cluster as `noise:coordinated-creation`
25
26 ## Evolution Expectation
27 Manipulation vectors rotate weekly. When a new metric vector appears (beyond stars and
28 forks), add a detection rule here. Watch for: issue-count inflation, sponsor-badge
29 gaming, discussion-count manipulation.
30
31 ## Integration
32 - Pre-filter step before editorial scoring
33 - Flagged repos should be noted in analysis frontmatter for retrospective audit
34 - Noise counts feed the weekly brief's Signal/Noise section