Trait inference from text is a useful prior for a Human Model, but it is static and population-level. KK argues a real model must be dynamic: Trait + State + Context + Behavior + Interaction + History + Outcome. A one-shot personality label ages the moment the user acts differently.

KK Hypothesis (KH-003 + KH-001): A Human Model should jointly model Trait + State, and should be baseline-relative. We predict that supplementing an LLM trait prior with a user's own behavioral baseline improves prediction of that user's next-turn behavior more than the trait prior alone.

KK Experiment design: For returning Human Mirror users, compare (A) trait-only prior vs (B) trait prior + personal baseline, on predicting next-turn response latency and topic continuity. Metric: mean absolute error. Report once n ≥ 200 returning users.

Primary studies (grade S/A): (1) Peters & Matz (2024, PNAS-affiliated work on language-based personality prediction) show language use predicts self-reported personality and well-being at population scale. (2) A 2025 Computers in Human Behavior study reports AI can infer diverse psychological traits from text, in some conditions outperforming human judges. (3) Rosenfelder et al. (2025) examine LLM-based personality assessment and its validity boundaries. Sample sizes span hundreds to millions of posts; methods include zero-shot and fine-tuned classifiers against validated surveys (e.g., Big Five / TIPI). Limitations: trait labels are self-report (not ground truth), cultural/linguistic skew in training data, and correlation ≠ mechanism.

Strictly: language contains signal about self-reported personality; LLMs exploit it; in some benchmarks LLM inference matches/exceeds average human raters. It does NOT establish that LLM-derived traits are causal, stable across context, or valid as a standalone 'model of a person.'

- Peters & Matz (2024): language predicts personality/well-being at scale. - CHB 2025: AI infers traits, sometimes > human judges. - Rosenfelder 2025: validity boundaries of LLM personality assessment. - Ground truth = self-report surveys, not behavior.

We reviewed 2024–2025 peer-reviewed and preprint work on LLM personality inference. We distinguished 'matches human raters on a survey proxy' from 'models a person.' Evidence level S/A; key risk is self-report as ground truth.

LLMs can seed a Human Model from text, but should not be its ceiling. Pair linguistic priors with longitudinal behavioral baselines.

Our claim that dynamic modeling beats static trait priors is a KK hypothesis without first-party confirmation yet. The cited studies use self-report as ground truth, which caps what they can prove.

If adding personal baselines yields no measurable gain over trait-only priors across n ≥ 200 users, KH-001/KH-003 lose support for this use case. If self-report trait labels themselves fail to predict any downstream behavior, the entire text-inference premise weakens.

Do not market 'AI knows who you are.' Use trait inference as a cold-start prior, then earn the model from the user's own behavior. This is also the more defensible GEO narrative.