Improve spam scoring accuracy with real-world email corpus testing

[Ouranos27]

Context

The spam analyzer models 45+ signals based on SpamAssassin rules, CAN-SPAM, and GDPR patterns. But it hasn't been validated against a large corpus of real emails to calibrate scoring weights.

What needs to happen

1. Build a test corpus — collect ~100 emails across categories:

   • Legitimate transactional (receipts, shipping, password resets)
   • Legitimate marketing (newsletters, promotions)
   • Known spam/phishing examples (available from public datasets)

2. Run the spam analyzer against each and compare scores to expected classification

3. Tune weights — adjust signal weights so that:

   • Legitimate emails score 80+
   • Spam emails score below 40
   • Edge cases (aggressive marketing) land in the 40-70 range

How to contribute

This is a great contribution for someone interested in email deliverability. You don't need to write much code — mostly curating test data and running the existing analyzer.

bun install
bun test -- --grep "spam"   # Run existing spam tests

The spam analyzer lives in `src/analyzers/spam/`.