TwoNN-intrinsic-dim regime boundary: Psi-vs-persona AUC-Delta drops by 0.05-0.15 per unit d_intrinsic in the (5,8] band
The 'curse of dimensionality' may degrade AI persona detection smoothly, not suddenly — and we can predict exactly how fast.
Curse-of-dim regime prediction sharpened from nominal to intrinsic dim axis (TwoNN); regime boundary tested as a slope (not a step), addressing Critic phase-transition-vs-continuous-degradation framing concern.
4 bridge concepts›
How this score is calculated ›How this score is calculated ▾
6-Dimension Weighted Scoring
Each hypothesis is scored across 6 dimensions by the Ranker agent, then verified by a 10-point Quality Gate rubric. A +0.5 bonus applies for hypotheses crossing 2+ disciplinary boundaries.
Is the connection unexplored in existing literature?
How concrete and detailed is the proposed mechanism?
How far apart are the connected disciplines?
Can this be verified with existing methods and data?
If true, how much would this change our understanding?
Are claims supported by retrievable published evidence?
Composite = weighted average of all 6 dimensions. Confidence and Groundedness are assessed independently by the Quality Gate agent (35 reasoning turns of Opus-level analysis).
Empirical Evidence
How EES is calculated ›How EES is calculated ▾
The Empirical Evidence Score measures independent real-world signals that converge with a hypothesis — not cited by the pipeline, but discovered through separate search.
Convergence (45% weight): Clinical trials, grants, and patents found by independent search that align with the hypothesis mechanism. Strong = direct mechanism match.
Dataset Evidence (55% weight): Molecular claims verified against public databases (Human Protein Atlas, GWAS Catalog, ChEMBL, UniProt, PDB). Confirmed = data matches the claim.
This hypothesis sits at the intersection of two somewhat obscure but important ideas: detecting subtle social signals in data, and a classic math problem called the 'curse of dimensionality.' The curse of dimensionality refers to a well-known headache in data science — as you add more dimensions (think: more variables describing each data point), the space becomes so vast and sparse that statistical methods start to break down. Density estimation, a technique used to figure out where data points cluster together, is particularly vulnerable. The more dimensions you have, the harder it becomes to find meaningful patterns. The hypothesis proposes something specific and testable: when AI systems try to distinguish between a person's genuine psychological profile (called 'Psi') versus a constructed 'persona,' their accuracy degrades in a predictable, gradual way as the true complexity of the data increases. The key word is 'true' — not the number of variables you feed the algorithm, but the real underlying complexity, measured using a clever technique called TwoNN that estimates what mathematicians call 'intrinsic dimensionality.' Think of it like this: a crumpled piece of paper exists in 3D space, but it's fundamentally a 2D surface. TwoNN finds that underlying '2D-ness.' The hypothesis then says: for every unit increase in this true complexity score within a specific range (roughly 5 to 8), the ability to tell a real profile from a fake one drops by a measurable, consistent amount — not a cliff-edge collapse, but a steady slope. What makes this interesting is the precision of the claim. It's not just 'things get worse' — it's 'things get worse at this specific rate, in this specific range, and here's the math explaining why.' That kind of quantitative prediction is rare and valuable because it's genuinely falsifiable.
This is an AI-generated summary. Read the full mechanism below for technical detail.
Why This Matters
If confirmed, this hypothesis could give researchers and engineers a practical early-warning tool: by measuring the intrinsic dimensionality of their data, they could predict in advance how reliable their AI-based persona or identity detection systems will be — before deploying them in high-stakes settings like fraud detection, social media authenticity checks, or psychological assessment tools. It could also shift how practitioners think about dimensionality reduction, encouraging them to measure true complexity rather than just count variables. The finding would suggest that there's no sudden cliff where systems fail, but rather a gradual, manageable degradation — which is actually more actionable, since it allows for calibrated confidence thresholds rather than binary pass/fail decisions. It's worth testing precisely because the quantitative slope prediction (0.05–0.15 AUC drop per unit intrinsic dimension) is specific enough to be proven wrong, which is exactly what good science requires.
Mechanism
AMISE-optimal bandwidth h_opt ~ n^{-1/(d+4)} and neighbour-count N_sphere = n V_d h_opt^d are functions of TRUE intrinsic dimensionality of kernel support (Silverman 1986; Wand-Jones 1995), not nominal target of UMAP projection. TwoNN estimator (Facco et al. 2017 Scientific Reports per Post-QG correction) recovers d_intrinsic from embedded coordinates directly. Substituting d_intrinsic into h_opt and N_sphere gives the actual decay-of-information curve. The hypothesis makes a SLOPE claim (severity smaller than phase transition): in d_intrinsic in [5,8] window, AUC-Delta = AUC(Psi) - AUC(persona) decays at rate 0.05-0.15 per unit d_intrinsic.
Supporting Evidence
AMISE-optimal bandwidth h_opt ~ n^{-1/(d+4)}, N_sphere = n V_d h_opt^d (Silverman 1986; Computational Validation Check 3). N_sphere ~ 68 at d=10, n=10^5 (Critic-re-derived). Abramson exponent -d/(2(d+4)) AMISE-optimal (Terrell & Scott 1992 Annals of Statistics). Per Post-QG Amendments: Facco et al. 2017 venue is Scientific Reports (not Nature Communications); Ansuini-BERT misattribution dropped, rely on per-panel TwoNN empirical measurement.
How to Test
For each panel (A=FOMC-day brokerage, B=CDC ZIP vacc): (1) UMAP at d_nominal in {2,4,6,8,10,15,20}; (2) TwoNN on embedded coords -> d_intrinsic; (3) recompute h_opt and N_sphere on d_intrinsic; (4) Psi-gradient detector on stance-blocked PD pilot, alpha=0.5, Abramson clip eps=0.01 of median; (5) elastic-net persona-logistic with inner 5-fold CV; (6) cluster-stratified outer 5-fold ROC-AUC for adoption inflection at 7d; (7) regress AUC-Delta on d_intrinsic with 1000-replicate panel bootstrap. Pre-register slope CI [-0.13, -0.03] on [5,8] window and Delta sign flip at d_intrinsic <= 5 vs >= 8.
Other hypotheses in this cluster
Asymptotic (1-AUC) floor model selection: Psi floor <= 0.10 vs Galesic/Jain-Singh floors >= 0.10/0.08 with crossing point n* in [10^4, 10^5]
A new mathematical benchmark could reveal which AI models for tracking public opinion are fundamentally limited — no matter how much data you feed them.
CSD/CSU on Psi-derived observables achieve 60-65% balanced accuracy at W=21d with continuous paid-spend label and explicit Poisson noise floor
Physics-borrowed 'tipping point' math may predict when social media buzz turns into real paid advertising.
Spectral-gap of audience-signal Laplacian predicts time-to-adoption-saturation: t_sat * gamma_2 in [0.7, 1.3] across panels
A single number from network math could predict how fast any market 'goes viral' — before it happens.
Two-tier conditional Psi advantage: Delta >= +0.08 at d_intrinsic <= 5 reverses to Delta <= -0.05 at d_intrinsic >= 8 with monotone interior gradient
Social media opinion signals may work well in simple debates but collapse in complex, high-dimensional ones.
Can you test this?
This hypothesis needs real scientists to validate or invalidate it. Both outcomes advance science.