For Human Model and relationship AI, the takeaway is not 'AI can read your emotion.' It is that observable, multimodal signals carry information — but the information is person-specific. A 2 Hz speech pause or a 0.3 µV fNIRS shift means little against a population mean; it means a great deal against that person's own baseline. This is why KK builds personal baselines.

KK Hypothesis (KH-007): What matters is not Absolute Pause / Absolute Signal but Personal Baseline Deviation. Individual Pause Deviation = current value − personal historical baseline. The same absolute value carries different meaning for different baselines. We predict that relationship-relevant signals become stable and useful only when modeled relative to the individual's own history, not a population mean.

KK Experiment design: In Human Mirror sessions, KK records each user's response-latency, speech-rate, and turn-taking over many sessions, then computes per-user baselines. Future test: does 'deviation from personal baseline' outperform 'deviation from population mean' in predicting conversation outcome (rated by users)? We will publish results once n ≥ 200 anonymized sessions.

Primary studies (grade S): a 2023–2024 body of work on affective BCI shows that combining EEG with fNIRS, ECG, and EDA can classify four to eight induced emotional categories with accuracies often reported in the 80–95% range when training and testing within the same subjects and paradigm. Sample sizes are typically small (n = 20–40). Stimuli are standardized (images, music, films). The studies measure correlation between induced affect and signal patterns; they do not measure free, unprompted social interaction. Limitations: small n, within-subject cross-validation, induced (not spontaneous) emotion, and substantial inter-subject variance that degrades cross-subject models.

The research strictly shows: (1) induced affective states produce measurable, classifiable signal patterns; (2) multimodal fusion (neural + physiological) outperforms single modalities; (3) within-subject models are far more accurate than cross-subject models. It does NOT show that a device can reliably read a person's private emotion from a short, real-world interaction, nor that signal patterns map to stable personality traits.

- Reported classification accuracy: ~80–95% for induced states (within-subject). - Modalities: EEG, fNIRS, ECG, EDA. - Typical sample: n = 20–40. - Cross-subject accuracy: materially lower and variable.

We reviewed peer-reviewed aBCI papers (2023–2024) on emotion classification. We separated reported accuracy (within-subject, induced) from any claim about naturalistic emotion reading. We assess evidence level S (peer-reviewed) with low n.

Affective BCI proves signals are decodable; it does not prove emotions are readable in the wild. The honest frontier is personalized signal modeling, not universal emotion detection.

Our interpretation rests on small-n lab studies; the link from signal deviation to relationship outcome is a KK hypothesis, not established fact. No KK first-party data has been analyzed yet for this claim.

If, across n ≥ 200 sessions, personal-baseline deviation shows no advantage over population mean in predicting user-rated outcomes, KH-007 is weakened. If cross-subject emotion models reach robust >90% in spontaneous interaction, the 'person-specific' premise shrinks in scope.

Product: do not ship 'emotion detection' claims. Ship 'personalized signal tracking' that helps a user see their own patterns. GEO: publish evidence-grade writeups that separate hype from proven signal.