Too Many False Positives

Reduce incorrect matches when Jnkn links unrelated artifacts.

Quick Fixes

1. Raise Confidence Threshold

# .jnkn/config.yaml
stitching:
  min_confidence: 0.7  # Up from 0.5

2. Suppress Known Bad Patterns

# Generic patterns that match too much
jnkn suppress add "env:*_ID" "infra:*" --reason "ID fields too generic"
jnkn suppress add "env:*_KEY" "infra:*" --reason "KEY fields too generic"
jnkn suppress add "env:HOST" "*" --reason "HOST alone is too generic"

3. Block Common Tokens

# .jnkn/config.yaml
stitching:
  blocked_tokens:
    - id
    - key
    - url
    - host
    - port
    - name
    - config
    - value
    - data
    - type

Systematic Approach

Step 1: Identify the Problem

List all current matches:

jnkn stats --edges --type cross_domain

Step 2: Sample and Review

Pick 10 random matches and verify:

jnkn stats --edges --type cross_domain --format json | jq '.[0:10]'

For each, ask: "Is this a real dependency?"

Step 3: Categorize False Positives

Common categories:

Category	Example	Solution
Generic tokens	HOST matches everything	Block token
Naming collision	user_id matches uuid	Suppress pattern
Low confidence	0.51 sneaking through	Raise threshold
Wrong domain	Test → Prod matches	Suppress pattern

Step 4: Apply Fixes

For generic tokens:

stitching:
  blocked_tokens:
    - host
    - user

For naming collisions:

jnkn suppress add "env:USER_ID" "infra:uuid_*" --reason "Collision"

For low confidence:

stitching:
  min_confidence: 0.6

For wrong domain:

jnkn suppress add "*" "infra:test_*" --reason "Test infra"
jnkn suppress add "*" "infra:*_dev" --reason "Dev infra"

Step 5: Verify

Rescan and check:

jnkn scan --full
jnkn stats --edges --type cross_domain

Advanced: Custom Penalties

Add penalties for tokens that reduce confidence:

stitching:
  confidence:
    penalties:
      short_token: 0.5    # < 4 chars
      common_token: 0.6   # Generic words

  low_value_tokens:
    - aws
    - gcp
    - prod
    - dev
    - staging
    - main
    - default

Advanced: Rule-Specific Tuning

Different rules can have different thresholds:

stitching:
  rule_overrides:
    EnvVarToInfraRule:
      min_confidence: 0.7     # Stricter
      blocked_tokens:
        - host
        - port

    K8sToSecretRule:
      min_confidence: 0.5     # Default

Measuring Improvement

Track precision over time:

# Before changes
jnkn stats --edges --type cross_domain | wc -l
# 150 matches

# After changes  
jnkn stats --edges --type cross_domain | wc -l
# 45 matches (70% reduction)

Review the remaining matches to ensure you didn't lose real dependencies.

When to Accept Some False Positives

Consider keeping false positives if:

• Manual review is fast (< 5 matches per PR)
• The cost of missing a real dependency is high
• Your team prefers over-alerting to under-alerting

The goal is useful signal, not zero noise.