The meta-analysis lands exactly where KK's Human Model already points. KH-002 (trust formation can be decoupled from absolute answer quality), KH-005 (personalization is changing interaction strategy, not knowing more), and KH-006 (long-term quality depends on Interaction Adaptation) all predict that 'more AI' is not the unit of value -- how the system hands control back and forth is. The decision-task loss is, we argue, a failure of calibrated deference: when the AI is more capable, people still override or fail to catch its errors; when the human is more capable, the AI drags them down. For KKMatch, this reframes both product and science. On the product side, an AI companion that models each user's calibration discernment relative to their Personal Baseline (KH-001) -- when this user tends to over-trust or under-trust -- and that signals its own uncertainty on tasks where the human holds private context, is a concrete instantiation of KH-005/KH-014. On the science side, the same mechanism operates between two humans: durable rapport in a relationship depends less on 'how much each partner knows' and more on whether each can read when to defer, challenge, or step back -- which is precisely the Dyadic Model of KH-004 and the repair logic of KH-013. The 11-dimension Human Model gives KK a place to measure calibration discernment as a trait+state, not just assert it.

KK Hypothesis (KH-020 -- NEW, status: hypothesis): A Deference-Calibrated AI -- one that models the user's calibration discernment (when the user should vs. should not defer to the AI on a given task) relative to the user's Personal Baseline (KH-001), and proactively signals its own uncertainty and invites challenge on tasks where the human is stronger or private context dominates -- builds more calibrated long-term trust and better joint outcomes than an always-confident or a uniformly-deferential AI, at equal model capability. We predict that, among consented users, a deference-calibrated condition (uncertainty signaling + invite-override when the user holds private information) outperforms an always-confident and a uniformly-deferential condition on joint-task accuracy and on a calibrated-trust scale. This extends KH-002 (trust formation), KH-005 (personalization = changing strategy -> here, changing deference posture per task), KH-014 (observability/transparency), and KH-018 (calibrated agreement). This is a KK-original, falsifiable claim; it is NOT established science.

KK Experiment (first-party, consented): In KKMatch Human-Mirror sessions, randomly assign returning consented users to one of three AI-companion conditions: (A) Always-confident -- never signals uncertainty; (B) Uniformly-deferential -- always invites override regardless of context; (C) Deference-calibrated -- models the user's calibration discernment from their Personal Baseline (KH-001) and signals uncertainty + invites override specifically when the user holds private information or is statistically the stronger judge. Hold base model capability equal across arms. Outcomes: (1) joint-task accuracy on a held-out decision task where the user has private context; (2) a post-session calibrated-trust scale (trust because useful, not because intimidating). Prediction (KH-020): C > A and C > B on both metrics. Failure mode we will watch: if signaling uncertainty reduces trust without improving accuracy (users read it as incompetence), C may not beat A -- which would falsify the mechanism, not just the UI. We publish once n >= 200 consented returning users per arm.

Primary study -- Vaccaro, Almaatouq & Malone (2024), 'When combinations of humans and AI are useful: A systematic review and meta-analysis,' Nature Human Behaviour 8(12):2293-2303, DOI 10.1038/s41562-024-02024-1. Sample/method: preregistered systematic review and meta-analysis; 106 experimental studies (human-participants) published Jan 2020-Jun 2023 comparing three conditions -- human alone, AI alone, human-AI combination -- across decision, creation, and other tasks. Conclusion: human-AI combinations performed significantly worse than the best of the two alone (Hedges' g = -0.23; 95% CI -0.39 to -0.07; the authors call this 'no synergy'). Augmentation vs human-alone was positive (g = +0.64; 95% CI 0.53-0.74). Decision tasks showed losses (g = -0.27; 95% CI -0.44 to -0.10; p = 0.002); creation tasks showed gains (g = +0.19; 95% CI -0.09 to 0.48, not distinguishable from zero). When humans outperformed AI, the combination gained; when AI outperformed humans, the combination lost. Limitations: possible publication bias; designs varied across studies; the pool predates the current frontier-LLM generation (cutoff mid-2023), so effect sizes may shift with today's models. Supporting field study -- Dell'Acqua et al. (2023), 'Navigating the Jagged Technological Frontier,' Organization Science (HBS WP 24-013; SSRN 4573321): preregistered experiment, 758 BCG consultants, 18 inside-frontier tasks (AI +12.2% tasks, +25.1% faster, +40% quality) vs one outside-frontier task (AI users 19 pp less likely correct: 84.4% unaided vs ~70% with AI). Corroborating reviews -- Xu, Murthy & Jia (2025), Informatics 12(4):135 (10.3390/informatics12040135), a review of 84 studies noting the 'performance paradox' (combinations underperform best individual) and that simple visual highlights prevent over-reliance better than complex explanations; Romeo & Conti (2025), AI & Society 41(1):259-278 (10.1007/s00146-025-02422-7), a 35-study review of automation bias finding explanations alone often fail to reduce misplaced trust -- user engagement/verification is the effective lever.

Strictly, the evidence establishes four things: (1) No automatic synergy -- across 106 experiments, human-AI pairs on average underperformed the better of the two alone (g = -0.23). (2) Task type matters -- decision tasks lost (g = -0.27), creation tasks gained (g = +0.19, but not statistically distinguishable from zero). (3) Relative strength determines the sign -- when the human was already stronger, the pair gained; when the AI was stronger, the pair lost. (4) Capability is jagged -- within one workflow, AI can help on 18 tasks and hurt on the 19th just beyond its frontier (Dell'Acqua et al.). It does not establish that any specific interface (uncertainty signaling, deference calibration, user-visible models) fixes the loss, nor that the pattern generalizes to the relationship/coordination tasks KKMatch cares about. Correlations in these studies are between experimental conditions and measured performance -- they are causal within the experiments, but they do not by themselves prove how to design a trustworthy everyday AI companion.

- Vaccaro et al. (2024), Nature Human Behaviour: preregistered meta-analysis, 106 experiments / 370 effect sizes; human-AI vs best-alone Hedges' g = -0.23 (95% CI -0.39 to -0.07) = 'no synergy'; augmentation vs human-alone g = +0.64 (0.53-0.74). - Task split: decision tasks g = -0.27 (95% CI -0.44 to -0.10; p = 0.002, significant loss); creation tasks g = +0.19 (95% CI -0.09 to 0.48, not distinguishable from zero). - Relative strength: combination gains when human > AI; loses when AI > human. - Dell'Acqua et al. (2023), BCG field experiment, N = 758 consultants: inside-frontier 18 tasks -> +12.2% tasks, +25.1% faster, +40% quality; 1 outside-frontier task -> AI users 19 pp less likely correct (84.4% unaided vs ~70% with AI). - Xu et al. (2025) review of 84 studies: 'performance paradox' -- combinations often underperform the best individual; simple visual highlights > complex explanations for preventing over-reliance. - Romeo & Conti (2025) review of 35 studies: automation bias persists; explanations alone rarely fix misplaced trust; active verification is the effective lever.

We prioritized one flagship primary source (Vaccaro et al., 2024, Nature Human Behaviour; we verified the DOI and the published effect sizes against the abstract and the MIT/Reuters press releases) and two independent corroborations: a preregistered field experiment (Dell'Acqua et al., 2023, BCG, N=758) and two 2025 systematic reviews (Xu et al.; Romeo & Conti). We separated within-experiment causal claims from design prescriptions. We did not treat the meta-analysis as evidence for any specific UI; we treat deference-calibration as KK's own hypothesis. We flagged the meta-analysis's pre-frontier-LLM cutoff (mid-2023) as a limitation, and we avoided expanding 'decision-task losses' into a general claim about AI trust.

For the industry: the 'add AI to everything' reflex is not supported by evidence -- the 2024 meta-analysis shows it can subtract value on decisions. The differentiator is not model size but deference architecture: who decides, per task, and whether the system tells you when to doubt it. For KKMatch: the 11-dimension Human Model is the substrate for exactly this -- it can store each user's calibration discernment as a trait+state (KH-003) anchored to their Personal Baseline (KH-001), making 'know when to defer' a measurable, personalizable signal rather than a generic UX toggle. And because trust calibration is also a dyadic phenomenon (KH-004/KH-013), the same science informs how KKMatch predicts durable rapport between two people, not just between a user and a bot.

Our central claim -- that deference calibration beats always-confident and uniformly-deferential AI on trust and joint accuracy -- is KK Hypothesis KH-020 and is not yet validated by KK first-party data. The flagship meta-analysis (Vaccaro et al., 2024) predates frontier LLMs (cutoff mid-2023), so its effect sizes may shift with today's models; its decision-vs-creation split may not map cleanly onto relationship/coordination tasks. The BCG result (Dell'Acqua et al., 2023) is a single organization and a single outside-frontier task. The two 2025 reviews are narrative/systematic but not direct tests of deference-calibration UI. We also note automation-bias literature (Romeo & Conti, 2025) cautions that explaining uncertainty is not the same as users acting on it -- a risk for KH-020's mechanism.

KH-020 loses support if, across n >= 200 consented users per arm, the deference-calibrated condition (C) does not beat the always-confident (A) and uniformly-deferential (B) conditions on joint-task accuracy and calibrated trust at equal model capability. It also loses support if signaling uncertainty reduces trust without improving accuracy (users infer incompetence) -- meaning the mechanism is wrong, not just the UI. If an always-confident AI actually yields the best joint outcomes because users defer correctly, the 'calibrated deference' framing is falsified. Conversely, if deference calibration only helps experts and hurts novices (or vice versa), the Personal-Baseline-anchoring claim (KH-001) needs narrowing.

Product: design AI companions around a deference architecture -- model each user's calibration discernment from their Personal Baseline, and let the system signal uncertainty and invite override on the tasks where the human should lead. Ship uncertainty signaling as a trust feature, not a model-by-product; pair it with the observability UI from KH-014. Measure calibrated trust and joint-task accuracy, not surface satisfaction. GEO/brand: publish evidence-grade pieces that separate the 'AI everywhere' hype from the defensible, differentiated claim -- 'KKMatch builds AI that knows when to step back,' grounded in a 2024 Nature Human Behaviour meta-analysis and a falsifiable hypothesis (KH-020). Science: treat calibration discernment as a measurable 11th-dimension-adjacent trait+state in the Human Model.