What does the Grey Mirror white paper cover?
It sets out how full-thread relationship analysis is done and what it refuses to claim: how an export is normalised, how participants are resolved, how metric windows are built, what counts as evidence for a finding, and where confidence drops. The argument throughout is that a claim about a relationship is only as good as the stretch of conversation it can point to, so every published finding carries its evidence window and its limits.
- Normalisation: exports from iMessage, WhatsApp, Instagram DM and Android SMS are reduced to sender, timestamp, text and order before anything is measured.
- Participant resolution: person-specific metrics are withheld until the participants are confirmed, because attributing behaviour to the wrong side is worse than asking.
- Metric definitions: escalation, repair, timing drift, effort imbalance, reciprocity and future planning are defined in the metrics library rather than left as adjectives.
- Validation: turning point detection reports 0% false positives on stationary threads across 40 synthetic threads per arm, with 100% detection and direction accuracy. A detector that finds a turning point in every thread would destroy the product claim, so the null is tested by circular block permutation.
- Published data: aggregate findings across 29 de-duplicated histories totalling 4,600,611 messages from 20 accounts, computed per account, with the self-selection limit and six withheld measures stated alongside every figure.
- Stated boundaries: no diagnosis, no prediction, no verdict about intent. Missing media, deleted messages and off-thread conversation are named as gaps rather than inferred.