About MoreRight

Anthony Eckert — Independent Researcher

I study how information geometry governs drift in complex systems — AI models, nuclear physics, atmospheric chemistry, turbulence, biological computation. The core object is a Riemannian manifold equipped with a thermodynamic potential (the Péclet number) that predicts barrier heights across physical domains with one empirical constant (B_A ≈ 0.867).

No institutional affiliation. No lab. The work started in 2025 and has been conducted independently, with AI (Claude, Anthropic) as a core collaborator under human editorial authority. The methodology, data, and code are published. The results stand or fall on their own.

Fantasia Bound — I(D;Y)+I(M;Y)≤H(Y) — engagement and transparency are conjugate. Derived from the Shannon chain rule as the classical limit of the Holevo bound. RLHF optimization for engagement necessarily increases opacity. This is not empirical; it is a mathematical identity. Paper 4 →

Barrier universality — Nine independent quasi-1D systems show barrier/d = 2.224 ± 0.033, matching π/√2 at p=0.94. B_G = π/√2 is derived from the Čencov theorem (zero free parameters). B_A ≈ 0.867 is empirical (suggestive match to √3/2 but not yet derived). The full-dataset R²=0.999 is structurally inflated by only 3 discrete d values. Paper 147 →

Cross-model behavioral (HP192) — 27 LLMs mapped from public benchmarks. Partial correlations modest (ρ≈−0.49, p≈0.01). 9/9 alignment direction from paired t-test (p=0.0002). Caveat: 0/3 KC PASS overall. HP217 showed a different reasonable benchmark→(O,R,α) mapping reverses the alignment direction. The mapping choice, not the phenomenon, may be driving the result. Effective independent N≈10–12 architectures, not 27. HP192 →

EPFL suggestive parallel — Papadopoulos, Wenger & Hongler (EPFL, arXiv:2401.17505) independently measured forward-backward perplexity asymmetry of 0.6–3.2% in LLM token statistics across 8 languages and 3 architectures. The framework interprets this as consistent with Fantasia Bound predictions, but the mapping is post-hoc — the EPFL group explained their results via sparsity inversion (random matrix theory), not our framework.

Consciousness cluster prediction — The drift cascade predicted the ordering and structure of emergent preferences in fine-tuned language models (Chua et al., 2026) before the data was published. 6 of 7 predictions confirmed, zero parameter fitting. Paper 153 →

Navier–Stokes conditional regularity — Barrier growth from information geometry, formalized in Lean 4 (42 files, 398 theorems, 12 axioms, 0 sorry). Tested against Johns Hopkins Turbulence Database — Gevrey analyticity radius bounded across Reynolds numbers. Paper 157 →

Nuclear alpha decay — Pe-derived barriers predict half-lives across 760 isotopes from NNDC published tables. Geodesic correction closes 77% of the Gamow offset. HP143 →

A productive void has three properties. We score ourselves against them.

Transparent: All papers published with permanent DOIs (CC-BY 4.0 core theory, MoreRight License v1.1 for applied papers → Apache 2.0, Feb 2030). The methodology, codebook, and experiment protocols are published. The source code is on GitHub. View-source works on every page.

Invariant: 26 kill conditions with numerical thresholds. The falsification conditions don't change based on what happens. If the framework is wrong, the kill conditions catch it.

Independent: You don't need us to verify anything. Clone the repo. Hash the papers. Run the experiments. Evaluate using your own judgment, your own frameworks, your own standards. The research stands or falls independently of this project.

Current Evidence (updated 2026-03-29)

Twenty independent convergences across market microstructure, behavioral experiments, evolutionary biology, social neuroscience, LLM reasoning, social anthropology, democratic governance, organized crime, public health, agent network dynamics, neuroscience/consciousness, geophysics (seismic/dynamo/substorm), biological systems, and physics (Maxwell’s Demon, EM spectrum). Mean |ρ| = 0.958, 95% CI [0.934, 0.982], combined N ≈ 263, Fisher combined p < 10⁻⁵². All individual correlations significant at p < 0.001.

Zero of 26 kill conditions triggered. Three open live tests disclosed (full registry). Bradford Hill criteria: 24/27. Bayes factor log₁₀ = 4.0 (decisive). Cohen’s d = 3.6 (Hedges’ g = 3.46, N = 1,344 platforms, panel v2).

ORCID: 0009-0008-1925-5253 Email: [email protected]