How This Ranking Was Learned
This report is built from mailbox headers only. It never uses email bodies or attachments. Each row starts as a header-derived contact profile: direction, date, directness, subject line features, relationship duration, active months, weekend/evening timing, and whether the contact looks like a person, an automated sender, a service provider, or a professional/institutional relationship.
The scoring moved through an iterative correction loop. The first version overvalued raw interaction volume and mutual replies. The next version added personal-closeness signals, weekend/evening activity, professional/service penalties, and local calibration for known examples. The current version adds evidence confidence: a contact must have enough independent evidence before it can rank as a truly close relationship.
The important lesson was that friendship is not just frequency. Strong relationships tend to show durable, varied, direct, reciprocal communication over many months or years. False positives often show a short burst, a single recurring topic, property/service language, school/institutional context, or work/project vocabulary.
- High scores require enough evidence: substantial message count, active months, relationship span, and diverse subject roots.
- Weekend and evening messages help only when the relationship is otherwise sustained and non-service-like.
- Professional, tenant/property, engineering, institutional, transactional, and automated patterns are pushed away from the personal-closeness ranking.
- Generic warmth words such as hello, thanks, and checking in are weak evidence unless paired with richer personal/life/social signals.
- User feedback is treated as calibration data for cases headers cannot infer, such as family obligation, dormant friendship, business-friend overlap, or known inner-circle relationships.