The research validates the core of KK's design since KH-001 and KH-003: a Human Model built only from a one-time trait quiz is, by definition, blind to the part of you that actually shows up in a conversation. Wagner & Gander show the same person differs day to day; Lee & Beck show a person is a set of momentary profiles, selected by situation. For KKMatch, this means the 11-dimension Human Model must store two things: a Trait component (your stable center, from aggregated history) and a State trajectory (where you are right now, relative to your Personal Baseline). Reading someone's Tuesday-morning state as if it were their fixed self is exactly the error a Trait-only model makes. This is KH-005 (personalization = changing strategy): the system should adapt now to your current state, not to a static label. It also extends KH-007 / KH-011 (pause / prosodic baselines) — those are State signals measured relative to baseline, the same logic at the behavioral and acoustic level.

KK Hypothesis (KH-015, extending KH-001, KH-003, KH-005): A Dynamic Human Model — a user's stable Trait component plus a continuously-updated, baseline-relative State trajectory — predicts a person's next in-session behavior, expressed need, and felt rapport better than a static Trait-only profile. We predict that, among consented KKMatch users, models that use State-deviation-from-Personal-Baseline features outperform Trait-only features on next-turn engagement and on same-session rapport, especially in the first ~5 sessions before Trait estimates stabilize. This is a KK-original, falsifiable claim; it is NOT established science. The psychology literature proves personality is dynamic — it does not prove a dynamic model improves matching. That step is ours.

KK Experiment design (first-party, consented): In Human Mirror and matching sessions, for each consented user maintain both a Trait vector (updated slowly from aggregated history) and a State vector (updated each session, expressed as deviation from that user's Personal Baseline, KH-001). At prediction time, run an A/B/C where the matching / response model uses (A) Trait-only, (B) Trait + raw recent signals, (C) Trait + baseline-relative State deviation. Hold user history equal across arms. Outcome metrics: next-turn engagement (did the user continue / deepen the exchange) and a same-session rapport scale. Prediction (KH-015): C > A on both metrics, and C > B in early sessions (turns 1-5) where raw signals are noisier than baseline-relative deviations. We will publish results once n >= 200 consented users per arm.

Three real, current sources: (1) Wagner & Gander (2024/2025), 'Character strength traits, states, and emotional well-being: A daily diary study' — Journal of Personality (DOI 10.1111/jopy.12933). A two-week daily-diary study of N=199 German-speaking adults. They measured trait character strengths at baseline and daily character-strength states and affect. Result: trait and aggregated-state measures converged well, but most character strengths showed high within-person variability — the same person looked different from day to day. Variability was predicted by whether a strength was 'phasic' (situation-dependent) or 'tonic' (stable). (2) Lee & Beck (2025), 'Idiographic Momentary Profiles of Personality Facets' — Journal of Personality and Social Psychology (DOI 10.1037/pspp0000568). Experience-sampling data from N=245 undergraduates, 15,833 total surveys. They estimated person-specific clusters of Big Five facet states. Finding: most individuals express multiple distinct momentary profiles, profile distinctiveness reflects both level and pattern, and situation characteristics (the DIAMONDS taxonomy) predict which profile a person expresses. (3) Lee & Beck (2025), 'Idiographic Time-Varying Networks of Personality Facets' (PsyArXiv, osf.io/xa5v2) — intensive longitudinal data from N=392 (31,564 surveys). Applying time-varying network models, they found 76.8% of contemporaneous and 74.5% of lagged networks were non-stationary — the relationships among a person's traits changed across the 2-3 week sampling window. Limitations: samples are predominantly student / WEIRD, self-report, and observational (no causal claims about what drives the shifts).

Strictly, the evidence shows: (a) personality is not only between-person differences — within-person, day-to-day variation is large and real (Wagner & Gander); (b) a single person is better described by several momentary configurations than by one fixed point, and the situation a person is in helps determine which configuration shows up (Lee & Beck, JPSP); (c) the connections among a person's traits are themselves unstable over short windows — personality behaves like a dynamic system, not a fixed structure (Lee & Beck, time-varying networks). It does NOT show that any specific product architecture (Trait+State, baseline-relative, etc.) causes better matching or rapport, nor that day-to-day variation is itself the target rather than noise to average out. The studies measure personality dynamics, not relationship or matching outcomes.

- Wagner & Gander (2024/2025): N=199, 2-week daily diary; high within-person variability in most character strengths; trait↔aggregated-state convergence good; phasic strengths vary more than tonic. - Lee & Beck (2025, JPSP): N=245; 15,833 surveys; most individuals express multiple distinct momentary profiles; situation (DIAMONDS) predicts profile expression. - Lee & Beck (2025, time-varying networks): N=392; 31,564 surveys; 76.8% of contemporaneous and 74.5% of lagged trait networks were non-stationary across 2-3 weeks. - Stanford AI Index 2025: 78% of organizations reported using AI in 2024 (up from 55% the year before) — evidence AI is now embedded in daily life and work, raising the stakes for how rigorously systems model people.

We reviewed three peer-reviewed / open-preprint studies (Wagner & Gander 2024/2025; Lee & Beck 2025, JPSP; Lee & Beck 2025, PsyArXiv) and one authoritative industry report (Stanford AI Index 2025). All three psychology studies use experience sampling / daily diaries — the gold standard for capturing within-person variability — but all rely on self-report and convenience (mostly student) samples, so we treat the direction and existence of variability as established and the magnitude as sample-dependent. We did not equate 'personality is dynamic' with 'a dynamic model improves matching' — that link is KK's hypothesis, not the studies' finding. We separated proven dynamics from KK's product claim and flagged the student-sample limitation explicitly.

For the industry: shipping a 'personality quiz -> fixed profile' as the whole user model is scientifically outdated and practically blind — it freezes a moving target. The differentiator is a model that knows your center and tracks your movement around it. For KKMatch: the 11-dimension Human Model is exactly this — Trait (stable) + State (baseline-relative, live). It is what lets the system adapt in the moment (KH-005 / KH-006) and, via KH-010, carry a coherent self across sessions instead of re-deriving you from zero. A dynamic model is also the substrate for the compatibility claims in KH-004 / KH-012: you match on how two moving models coordinate, not on two static points.

Our central claim — that a baseline-relative Dynamic Human Model beats a Trait-only profile on engagement and rapport — is KK Hypothesis KH-015, without first-party confirmation yet. The three psychology studies are observational, self-report, and mostly student samples; their variability magnitudes may not transfer to a general dating / relationship population. They measure personality dynamics, not matching or relationship outcomes, so the leap to 'better matches' is ours. Causality (what drives the shifts) is unestablished; we are betting the baseline-relative framing is the right compression of the dynamics, but a flat State vector could prove sufficient or a different compression could win.

If, across n >= 200 consented users per arm, the Trait + baseline-relative State condition (C) does NOT beat Trait-only (A) on next-turn engagement and same-session rapport, KH-015 loses support — the dynamics may be noise to average out, not signal to track. If raw recent signals (B) match or beat baseline-relative State (C), then simple recency works as well as our baseline logic, weakening KH-001's centrality. If the effect appears only after many sessions (not in turns 1-5), the 'early-session advantage' clause of KH-015 is wrong. If relationship outcomes (not just in-session rapport) show no lift, the dynamic model helps conversation but not matching — bounding its value.

Product: build the Human Model as two stores — a slow Trait vector and a fast, baseline-relative State vector — and make State visible / editable to the user (extends KH-014's observability). Don't over-fit to a single session's raw signals; compress them as deviation from baseline. GEO / brand: publish evidence-grade writeups that separate 'you are a fixed Big Five score' (outdated) from 'you are a stable center plus a live, baseline-relative state' (what the 2024-2025 literature supports and what KK builds) — a differentiated, defensible narrative for KKMatch as a research-led relationship platform. Real case to watch: as AI assistants embed into daily life (78% of orgs in the AI Index 2025), the systems that model users as dynamic, baseline-relative entities will feel 'known' rather than 'filed.'