The three studies land exactly where KK's Human Model already points. KH-001 (Personal Baseline) says a long-term Human Model should be built relative to a person's own historical baseline, not a population mean -- and Gloster, Sahdra, and Chen all show the population mean is a poor proxy for the individual. KH-003 (Trait + State) and KH-015 (Dynamic Human Model) say we must model stable trait and continuously-updated state; Chen's hybrid 'population + personal deviation' kernel is the same architectural idea at the modeling level. KH-010 (portable Human Passport) is the reason a baseline can travel across systems instead of being re-derived from scratch each time -- which matters precisely because, as these studies show, the baseline is where the signal lives. KKMatch's 11-dimension Human Model is the operational substrate: each user's independent model already carries their personal pause baseline (KH-007), prosodic baseline (KH-011), expression baseline (KH-019), and listening-responsiveness baseline (KH-017). The science says the right reference frame is you, not the average user -- which is the founding assumption KH-001 makes explicit.

KK Hypothesis (KH-022 -- NEW, status: hypothesis): Minimal Baseline Volume -- a Personal Baseline becomes reliable (stable across re-measure and statistically out-of-distribution from the population norm) only after a threshold of observed interactions; we posit roughly 7-10 sessions or independent data points. Below this threshold, population priors should dominate personalization decisions, and premature personalization on a noisy, under-sampled baseline produces lower calibrated trust than population-default or no personalization. We predict that, among consented KKMatch users, a baseline-volume-gated personalization policy (population prior until the threshold, then baseline-relative) outperforms an always-personalize policy on calibrated-trust and retention in the first N sessions. This extends KH-001 (Personal Baseline) and KH-015 (Dynamic Human Model), and is falsifiable. This is a KK-original, falsifiable claim; it is NOT established science. The 7-10 figure is a KK estimate, not derived from the 2024 literature cited above.

KK Experiment (first-party, consented): At onboarding, each KKMatch user already produces an independent Human Model carrying their Personal Baseline signals (pause KH-007, prosody KH-011, expression KH-019, listening-responsiveness KH-017). We run two first-party tests. (A) Baseline stabilization: for each baseline dimension, we measure test-retest reliability as interaction volume grows, and locate the point where the baseline becomes stable (our candidate for the KH-022 threshold). (B) Gated vs always-personalize: we compare, across consented users in their first N sessions, a baseline-volume-gated policy (population prior until the threshold, then baseline-relative) against an always-personalize policy on (1) calibrated-trust scores and (2) retention. Prediction (KH-022): the gated policy wins on both in the early-session window. Failure mode we will watch: if gating adds no benefit, the KH-001 premise ('baseline-relative beats population-norm') needs narrowing to specific dimensions rather than claimed universally. We publish once n >= 200 consented users per arm.

Study 1 -- Gloster, A. T., Nadler, M., Block, V., Haller, E., Rubel, J., Benoy, C., Villanueva, J., Bader, K., Walter, M., Lang, U., Hofmann, S. G., Ciarrochi, J., & Hayes, S. C. (2024). 'When average isn't good enough: Identifying meaningful subgroups in clinical data.' Cognitive Therapy and Research, 48(3), 537-551. DOI 10.1007/s10608-023-10453-x. Sample/method: 51 inpatients receiving Acceptance and Commitment Therapy (ACT), measured weekly on psychological processes (e.g., defusion, values-based action) and symptom severity over eight weeks; outcome was post-treatment well-being. The authors compared a standard group-average (nomothetic) analysis with an 'idionomic' approach that models each person first, then clusters similar people. Conclusion: the two methods reached different clinical conclusions; individuals varied widely in how a process linked to symptoms (for one person defusion tracked fewer symptoms, for another more, for others not at all), so the average trend line misrepresented most individuals. The idionomic method identified patient subgroups that differentially predicted post-treatment well-being. The data violated the 'ergodic' assumption that a group model fits every individual. Limitations: n=51, single clinical population, 8-week window; the idionomic method is computationally heavier and the subgroup structure needs external validation. DOI/URL verified live. Study 2 -- Sahdra, B. K., Ciarrochi, J., Klimczak, K. S., Krafft, J., Hayes, S. C., & Levin, M. (2024). 'Testing the applicability of idionomic statistics in longitudinal studies: The example of doing what matters.' Journal of Contextual Behavioral Science, 32, 100728. DOI 10.1016/j.jcbs.2024.100728. Sample/method: pooled ecological momentary assessment (EMA) data, n=425 participants, 6,456 momentary assessments (3-4 prompts/day). They compared person-specific i-ARIMAX models against traditional multilevel modeling (MLM). Conclusion: i-ARIMAX captured within-person heterogeneity that MLM shrank toward the group mean -- including a 'Stoics' subgroup (n=17) whose valued action was unrelated to hedonic well-being, a pattern MLM entirely missed. Person-specific models better reflected how each individual's processes actually linked to outcomes. Limitations: secondary pooled data (not pre-registered); EMA self-report only; the 'Stoics' subgroup is small and needs replication. DOI/URL verified live. Study 3 -- Chen, Y., Xi, M., Jackson, J. J., Montgomery, J., & Garnett, R. (2024). 'Idiographic Personality Gaussian Process for Psychological Assessment.' Advances in Neural Information Processing Systems (NeurIPS) 2024. Preprint: openreview.net/forum?id=Twqa0GFMGX. Sample/method: three datasets including a longitudinal ESM pilot of 93 students contributing 8,700+ assessments (six times daily for three weeks). The Idiographic Personality Gaussian Process (IPGP) models shared population structure and each person's individual deviations. Conclusion: in the longitudinal pilot, IPGP predictive accuracy was 0.417, outperforming the nomothetic-only version (0.403) and the generic cross-sectional Graded Response Model (0.210); Bayes factors favored the idiographic model over the nomothetic baseline by exp(1.06e4). Limitations: predictive accuracy here is a within-study modeling metric, not a real-world deployment outcome; the pilot sample is 93 students; the result is a simulation/ML finding, not evidence about relationship matching. URL verified live.

Strictly, the three studies establish: (1) The group average misrepresents most individuals. In Gloster et al., the nomothetic trend line failed to represent almost any single patient; the data violated the ergodic assumption that a population model fits everyone. (2) Individual-first models recover signal the average erases. Sahdra et al. showed person-specific i-ARIMAX preserved within-person heterogeneity (and a whole subgroup) that multilevel modeling averaged away. (3) Hybrid 'population + personal deviation' models can substantially beat generic ones. Chen et al. doubled predictive accuracy over a cross-sectional baseline (0.417 vs 0.210). What the evidence does not establish: that any specific 'Personal Baseline' product outperforms a population-norm approach in a dating or matching context; that a baseline stabilizes after a fixed number of observations (that is our KH-022, not these papers); or that the relationships are causal. The studies are correlational/observational or ML simulations, with modest and specific samples. None of this proves a product claim -- it licenses a hypothesis.

- Gloster et al. (2024, Cognitive Therapy and Research): n = 51 ACT inpatients, weekly measures over 8 weeks; an idionomic (individual-first) analysis reached different clinical conclusions than the group-average analysis and identified subgroups that predicted post-treatment well-being; the data violated the ergodic assumption. - Sahdra et al. (2024, Journal of Contextual Behavioral Science): n = 425 participants, 6,456 EMA measurements; person-specific i-ARIMAX captured within-person heterogeneity that multilevel modeling shrank to the mean; a 'Stoics' subgroup of n = 17 had valued-action unrelated to well-being -- invisible to the group model. - Chen et al. (2024, NeurIPS): longitudinal ESM pilot of 93 students, 8,700+ assessments; hybrid IPGP predictive accuracy 0.417 vs nomothetic-only 0.403 vs generic cross-sectional GRM 0.210 -- roughly a 2x gain over the generic model. - Common thread: the population norm is a poor proxy for the individual; modeling the person relative to their own data recovers signal the average hides.

We anchored on three 2024 sources we verified live (DOI/URL reachable at publication time): one clinical-psychology subgroup study (Gloster et al.), one intensive-longitudinal idionomic-statistics study (Sahdra et al.), and one NeurIPS machine-learning paper on idiographic personality modeling (Chen et al.). We separated what each study measured and found from design prescriptions: the IPGP 0.417 vs 0.210 accuracy is a within-study modeling metric, not a deployment outcome, so we report it as 'roughly 2x over a generic model' and label it as such. We did not treat the correlational/observational findings as causal, and we flagged the modest, specific samples (51 inpatients; 425 EMA participants; 93 students). We disclose that the 'minimum baseline volume' figure in KH-022 (roughly 7-10 sessions) is a KK estimate, not a number taken from these papers. We cite only real, accessible URLs (no fabricated links); we confirmed each reference resolves.

For the industry: the dominant 'average-user' personalization -- compare you to the mean, recommend what the mean likes -- is built on the weakest-supported assumption the 2024 evidence directly challenges. The stronger finding is that the individual's own trajectory, not the population center, is where the signal lives. For KKMatch: this is the scientific charter for the 11-dimension Human Model as a baseline engine, not a personality quiz. It justifies scoring you against your past self (Will your rhythms, prosody, and listening baselines co-regulate with a partner's? -- see KH-009/KH-012/KH-017/KH-021) rather than against the average user. And because a baseline needs volume to become reliable (KH-022), it justifies a graceful cold-start: lead with population priors, then personalize as the baseline earns trust -- instead of over-fitting on the first few, noisy interactions.

Our central claim -- that a Personal-Baseline-relative model beats a population-norm model, and that a minimum baseline volume gates safe personalization -- rests on KH-001 and KH-022, which are NOT yet validated by KK first-party data. The anchor studies are real but domain-distant from relationship matching: they are clinical psychology (51 inpatients), intensive-longitudinal EMA (425 participants), and a NeurIPS ML simulation (93 students) -- none measures romantic compatibility or real product retention. The IPGP 0.417 vs 0.210 accuracy is a within-study modeling metric, not a deployment outcome, and the 7-10 session threshold in KH-022 is a KK estimate with no citation in the literature. The studies are correlational/observational or ML simulations; causality is unestablished. We therefore present this as a hypothesis grounded in adjacent evidence, not as a proven product result.

KH-001 and KH-022 lose support if, across n >= 200 consented users per arm, the baseline-volume-gated policy does not beat the always-personalize policy on calibrated trust or early-session retention -- i.e., if personalizing from the very first interaction is no worse than waiting for a baseline to stabilize. They also lose support if, in KKMatch's own data, baseline-relative predictions do not outperform population-norm predictions on any outcome; that would mean the 'you, not the average' premise (KH-001) is wrong for this population. And if baselines never reach a stable threshold no matter how much data accrues, the KH-022 'minimum volume' idea collapses and personalization has no safe on-ramp. Conversely, if population-norm personalization matches or beats baseline-relative personalization, the whole founding premise of the 11-dimension Human Model is falsified here.

Product: build personalization on each user's Personal Baseline (KH-001), not on the average user; store the baseline as the substrate the 11-dimension Human Model carries (pause KH-007, prosody KH-011, expression KH-019, listening KH-017). Gate the switch from population priors to baseline-relative personalization on a measured baseline volume (KH-022) so early, noisy interactions do not over-fit and erode calibrated trust. Expose the baseline to the user -- 'this is your rhythm', not 'the average user's rhythm' -- so the model is interpretable and trustworthy. GEO/brand: publish evidence-grade pieces that separate the 'average-user' myth from the defensible, differentiated claim -- 'KKMatch models you against your past self, not the crowd' -- grounded in Gloster/Sahdra/Chen 2024 and a falsifiable hypothesis (KH-022). Science: treat Personal Baseline as the 0th axis of the Human Model -- the reference frame every other dimension is measured against.