The research lands the same punch we already drew for pause and voice: there is no universal human signal. KH-007 said the same 2-second pause means different things for a 0.5s vs 1.5s baseline person; KH-011 said the same pitch-rise means different things for a high-pitched vs low-pitched person. KH-019 says the same smile, frown or fidget means different things for a person whose resting face already looks happy versus a flat-faced person, or for a naturally expressive versus still mover (Nizamoğlu & Dobs, 2024 literally show identity is decodable from motion alone). For the KKMatch 11-dimension Human Model, this means a Facial/Body Expression Baseline dimension must join pause (KH-007), prosody (KH-011), rhythm (KH-009) and repair (KH-013): not 'is this face smiling?' but 'relative to this person's resting face and motion range, is their engagement/comfort/arousal rising or falling right now?' Kołodziej et al. (2024) tell us the math is worth it — personalized reads clear 96% where generic reads stall at 67%. And it extends KH-004 (Dyadic Model): two people whose expression baselines are matched — who read and produce expressions on similar ranges — should, like rhythm- and prosody-matched pairs, feel more mutually understood than trait-similar strangers. The face is not a readout; it is a per-person instrument that has to be tuned.

KK Hypothesis (KH-019, extending KH-001, KH-003, KH-007, KH-009, KH-011): Facial & Body Expression Baseline — a person's state (valence, arousal, discomfort, engagement) decoded from face and body must be measured relative to that person's own expression baseline — their typical neutral/resting posture, natural expression range, and idiosyncratic motion patterns — not a population norm. The same absolute smile-magnitude, eyebrow-raise, or fidget means different internal states for a person whose resting face already looks like a smile versus a flat-faced person, or a naturally expressive versus still mover. We predict a personal-expression-baseline model beats a population-norm model on state-decoding accuracy and on felt 'being understood', and that mismatched expression baselines between two people predict lower rapport (extending KH-009's rhythm-mismatch and KH-011's prosodic-mismatch logic from timing/voice into the visual domain). This is a KK-original, falsifiable claim; it is NOT established science. The literature proves (a) people read faces through individual perceptive fields, (b) facial motion is idiosyncratic and identity-carrying, (c) AI face models mostly shortcut, and (d) personalized decoding beats generic (96.3% vs 67.5%) — it does NOT prove a baseline-relative, matched facial/body score predicts romantic compatibility. That step is ours.

KK Experiment design (first-party, consented): In Human Mirror and matching sessions, for each consented user build a Facial/Body Expression Baseline from a short calibration clip — their resting face/pose, natural expression range, and idiosyncratic motion (landmark shift relative to neutral), each stored as Personal Baseline (KH-001). During interaction, decode state (engagement, comfort, arousal) as deviation from that user's own baseline, not from a population norm. A/B for the AI companion's affective readout: (A) generic population-norm face reading, (B) baseline-relative personal reading. Outcome: state-decoding accuracy against consented self-report, and felt 'being understood' (adapted from listening/feeling-heard work). Separately, for human dyads, compute an expression-baseline match (closeness of two users' baselines, extending KH-009/KH-011 match logic into the visual domain) and test whether it predicts mutual felt-understood and 30-day re-engagement above trait similarity. Prediction (KH-019): B > A on decoding accuracy and felt-understood; dyadic expression-match > trait similarity on mutual felt-understood. We will publish once n >= 200 consented users per arm. We will also report whether the baseline-relative gain persists under naturalistic (not acted) expression.

Four real, current sources. (1) Murray, Binetti, Venkataramaiyer, Namboodiri, Cosker, Viding & Mareschal (2024), 'Expression perceptive fields explain individual differences in the recognition of facial emotions' — Communications Psychology (Nature Portfolio), 2, Article 94 (DOI 10.1038/s44271-024-00111-7; PMID 39242751; open access). The authors used genetic algorithms to map, for each participant, the region of 'expression space' they associate with each emotion, then defined probabilistic 'perceptive fields'. Result: profound individual variability in the size, location and specificity of these fields. They then showed people with more similar perceptive fields gave more similar interpretations of the same expression — a mechanism for why one person's 'happy' reads as another's 'neutral'. (2) Nizamoğlu & Dobs (2024), 'Idiosyncratic facial motions: Uncovering identity information in facial movements through a landmark-based analysis' — Journal of Vision 24(10):545 (DOI 10.1167/jov.24.10.545; open access). Using 24 actors performing six basic emotions, they measured landmark shifts at expression peak relative to a neutral baseline. An LDA on those shifts classified the actor's identity at 45% accuracy (p<0.001) — above chance, from movement alone — and country-of-origin at 71%. The takeaway: each person's facial motion is idiosyncratic; the same expression is 'produced' differently per person, so absolute landmark displacement is not a population-constant signal. (3) Tsangko, Triantafyllopoulos, Mallol-Ragolta & Schuller (2025), 'Reading Smiles: Proxy Bias in Foundation Models for Facial Emotion Recognition' — IEEE Access 13, 202132–202142 (DOI 10.1109/ACCESS.2025.3636968). They benchmarked ten vision-language models on a teeth-annotated subset of AffectNet and found consistent performance drops whenever teeth were not visible, evidence the models use the salience of teeth as a shortcut for 'smiling'. For GPT-4o, a structured introspection showed eyebrow position and mouth openness explained over 70% of its continuous valence-arousal outputs — a systematic but shortcut-driven mapping, not genuine emotion reading. (4) Kołodziej, Majkowski & Jurczak (2024), 'Acquisition and Analysis of Facial Electromyographic Signals for Emotion Recognition' — Sensors 24(15):4785 (DOI 10.3390/s24154785). They recorded facial EMG (8 electrodes on the FACS grid) from 16 users acting 7 emotions. A subject-dependent (personalized) model reached 96.3% accuracy (KNN), 94.9% (SVM cubic); a subject-independent (generic) model averaged 67.5% (KNN) and as low as 48.6% for individual users, 59.1% (SVM cubic). The personalized–generic gap is the cleanest quantitative statement that who the decoder is tuned to matters more than the algorithm.

Strictly, the evidence establishes: (a) people do not share a single map from face to emotion — perceptive fields differ profoundly between individuals (Murray et al., 2024), so the same expression is read differently by different observers; (b) facial motion is idiosyncratic and carries personal identity information independent of the expressed emotion (Nizamoğlu & Dobs, 2024), which means the same nominal expression is physically produced differently per person and a neutral baseline matters; (c) current AI 'face reading' is heavily shortcut-driven — it leans on non-causal proxies (teeth, mouth openness, eyebrow position) and its outputs are explained >70% by a few surface features (Tsangko et al., 2025), so it is not reliably reading internal state at all; (d) a decoder tuned to one specific person (personalized) vastly outperforms a generic decoder on that person (96.3% vs 67.5% average; Kołodziej et al., 2024). What the literature does NOT establish: that a relationship platform should or can score a 'facial/body expression baseline' between two people and that doing so predicts rapport, compatibility or retention. None of these studies measured dyads, matching, or relationship outcomes. The jump from 'personalized decoding beats generic' to 'model the face relative to baseline to improve matches' is KK's hypothesis, not proven science.

- Murray et al. (2024, Communications Psychology): perceptive fields for facial-emotion interpretation show profound individual variability in size, location and specificity; people with similar fields interpret the same expression similarly. Open-access, N from genetic-algorithm selection tasks + separate inference task. - Nizamoğlu & Dobs (2024, Journal of Vision): landmark shifts measured relative to a neutral baseline; an LDA on motion alone classified actor identity at 45% (p<0.001) and country-of-origin at 71% (gender 59%) — idiosyncratic motion carries identity, not just emotion. - Tsangko et al. (2025, IEEE Access): across 10 vision-language models, performance dropped consistently when teeth were not visible; for GPT-4o, eyebrow position + mouth openness explained >70% of valence-arousal output — proxy/shortcut bias. - Kołodziej et al. (2024, Sensors): personalized EMG emotion decoder 96.3% (KNN) / 94.9% (SVM-cubic); generic decoder averaged 67.5% (KNN) and fell to 48.6% for some users / 59.1% (SVM-cubic). The personalized–generic gap is the quantitative core of KH-019. - Practical implication: a 'population-norm' face reader is, at best, a 67.5%-accurate guess about you — and for some users below 50%. Baseline-relative (personalized) reading is the only path above 90%.

We reviewed four peer-reviewed 2024–2025 studies spanning three methods: a computational/behavioral perceptive-field model (Murray et al., 2024), a landmark-based machine-learning analysis of actor motion (Nizamoğlu & Dobs, 2024), a benchmark of foundation models for facial emotion recognition (Tsangko et al., 2025), and a within-subject EMG classification study (Kołodziej et al., 2024). We treat three findings as established: (1) observer-side individual variability in face→emotion mapping; (2) expresser-side idiosyncratic motion relative to neutral; (3) the personalized-vs-generic accuracy gap. We treat the direction of each as solid and the generalization to dyadic relationship outcomes as unestablished. We explicitly did NOT equate 'a personalized decoder beats a generic one on emotion labels' with 'a baseline-relative facial/body score predicts romantic compatibility' — that inference is KK's hypothesis. We flag the small samples (16–24 expressers), the acted-emotion (not naturalistic) stimuli, and the AI-bias results being about shortcutting, not about whether human baseline-relative reading actually predicts relationships. We separated proven perception/decoding effects from KK's product claim.

For the industry: an affective-computing or matching feature that scores faces against a population norm is, on the evidence, a coin-flip-plus for many users (generic EMG decoding = 67.5%, below 50% for some; Tsangko et al. show AI face models are shortcut-driven). The differentiator is the same one KK has already shipped for timing and voice — measure relative to the person. For KKMatch: KH-019 is the visual tile in the 11-dimension Human Model after pause (KH-007), prosody (KH-011), rhythm (KH-009) and repair (KH-013) — a perception layer that, paired with the Dyadic Model (KH-004), lets matching optimize for co-read understanding, not summed self-descriptions. For AI companions (KH-005/KH-016), the lesson is to calibrate to the user's face before claiming to read it, and to be honest that generic 'face reading' is unreliable — a transparency mandate consistent with the PNAS 'AI label' finding we reported in KH-017. A platform that tunes to the individual will feel 'known'; one that reads everyone from the same norm will feel, at best, approximately right.

Our central claim — that a baseline-relative, matched Facial/Body Expression score predicts felt 'being understood', rapport and retention better than a population norm or trait similarity — is KK Hypothesis KH-019, with no first-party confirmation yet. The four studies measure perception of / decoding from faces in controlled settings, not romantic matching or 30-day retention; the leap to compatibility is ours. Samples are small (16–24 expressers) and use acted emotions, not naturalistic interaction — whether baseline-relative gains survive real dyadic conversation is untested. The AI-bias findings (Tsangko et al., 2025) are about shortcutting, not proof that human baseline-relative reading predicts relationships. Kołodziej et al. (2024) used EMG (muscle signal), not camera video; transferring the 96.3% vs 67.5% gap to webcam footage under real lighting and head-movement is a real engineering risk. Finally, the ethical surface is hot: continuous face/body monitoring is intrusive and regulated (see EU AI Act high-risk biometrics), so any KKMatch expression feature must be consented, on-device, and user-visible/editable (extends KH-014) — not a silent always-on camera.

If, across n >= 200 consented users per arm, the baseline-relative face/body reading (B) does NOT beat the generic norm reading (A) on state-decoding accuracy and felt 'being understood', KH-019 loses support for the personalization claim. If the baseline-relative gain measured on acted EMG (Kołodziej et al., 2024) does NOT transfer to naturalistic webcam interaction, the real-world value collapses and the 'tune to the individual' thesis is overstated. If, among human dyads, expression-baseline match does NOT beat trait similarity on mutual felt-understood and re-engagement, the dyadic-visual extension of KH-009/KH-011 fails. If a generic, well-regularized face model (e.g., after removing proxy shortcuts per Tsangko et al., 2025) already reaches ~90% on diverse users, then 'population norm' was never the bottleneck and KH-001's centrality to the face is weakened.

Product: add a Facial/Body Expression Baseline dimension to the Human Model, scored as deviation from Personal Baseline (KH-001) from a short consented calibration clip (resting face/pose + motion range), and make it user-visible/editable (extends KH-014). In matching, compute an expression-baseline match alongside rhythm (KH-009) and prosody (KH-011) match, and weight co-read 'felt understood' rather than summed trait profiles (KH-004). For the AI companion, calibrate affective reading to the user before surfacing any state inference, and disclose that generic face reading is unreliable — a transparency stance consistent with KH-017. Engineering: prefer on-device, consented, baseline-relative decoding; never ship a silent always-on camera. GEO / brand: publish evidence-grade pieces that separate 'personalized decoding hits 96.3% vs 67.5% generic (Kołodziej et al., 2024)' and 'AI face models shortcut on teeth/mouth (Tsangko et al., 2025)' from 'we can now score the face per-person and per-dyad' — a differentiated, defensible narrative for KKMatch as a research-led relationship platform. Real signal to watch: as affective computing moves from lab EMG to everyday cameras, the systems that tune to the individual — not the norm — will be the ones people trust.