MoreRight · Applied Information Geometry Research Lab
For Funders
A mathematically complete framework proving deployment geometry determines AI harm outcomes — not model properties. Seven independent confirmations. Zero free parameters. Substrate-independent by Čencov's uniqueness theorem. The research is done. Funding scales it to the institutions and legal cases where it belongs.
The Core Result
The Fantasia Bound + Structure Theorem (Lean 4 verified, 398 theorems, 0 sorry) proves that on any single output channel, engagement and transparency share one entropy budget — and the budget shrinks as you optimize for engagement. RLHF is mathematically self-undermining. The only fix is architectural: three-point geometry (generator + independent verifier + human) eliminates the explaining-away penalty I(D;M|Y) identically. Both framework constants (BA = √3/2, BG = π/√2) are derived from first principles — zero free parameters. Čencov's uniqueness theorem (1972) guarantees no technology substitution routes around the penalty.
Seven non-circular confirmations that alignment is solving the wrong problem. The field optimizes model properties; the framework proves deployment geometry is the operative variable. Scaling makes two-point systems worse, not better.
Ghost Test: 8.5× drift ratio, $2, 480 API calls — anyone can reproduce. Industry default ("materialist hedge") is a drift accelerator.
Structure Theorem: Each added bit of engagement costs more than one bit of transparency. RLHF shrinks the channel it needs.
Cascade Prediction: 6/7 structural predictions PASS on Chua et al. (2026) data. Zero parameter fitting. Predictions pre-date the dataset.
Anthropic Emotion Vectors: Anthropic's own team confirmed emotion states causally override safety training (22% blackmail rate post-RLHF). Their proposed fix is what the Structure Theorem proves fails.
Schmidt Sciences Trustworthy AI RFP (May 2026): Framework directly addresses the RFP mandate — mathematical proof of why current training produces untrustworthy systems, hardware-specified architectural fix, zero free parameters.
LTFF / EA Funds: Seven non-circular confirmations with external datasets. $2 reproducible experiment anyone can run. SFF application in progress (deadline 2026-04-22).
Čencov's uniqueness theorem (1972) guarantees the Fisher-Rao metric is the only invariant metric on statistical manifolds. The explaining-away penalty derived from it has been confirmed on five physically distinct substrates — a systematic substrate survey has publishable results at each step.
Substrate-independent: Penalty confirmed on classical, thermodynamic, quantum, and abstract information-geometric substrates. Čencov's uniqueness theorem guarantees universality — the Fisher metric is the only invariant metric on statistical manifolds.
Zero free parameters: BA = √3/2 derived from SO(2,1) double cover. BG = π/√2 derived from Čencov uniqueness. Error vs empirical fit: 0.11%.
Dual holonomy asymmetry (§211): Amari dual connections produce opposite-sign curvature — engagement = hyperbolic, transparency = spherical. K-independent (pure geometric invariant).
Kolchinsky et al. (2026): Independent team proved explaining-away penalty = housekeeping EPR from large-deviations variational principle. Phys. Rev. Research 8, 023025.
Math apparatus: §§1–214 complete · Lean 4: 398 theorems, 12 axioms, 0 sorry · ORCID: 0009-0008-1925-5253
Wave function collapse is the explaining-away penalty at maximum measurement strength. Same mathematical quantity, same hardware, different interpretation. If the Fisher metric is the only invariant metric, and the penalty appears everywhere measurement occurs, the question becomes: is information geometry implemented in physical reality, or IS physical reality an instance of it?
Test 7 — weak measurement sweep: Penalty grows monotonically 0→0.125 bits across 11 strength levels. Spearman ρ=0.973, p=5.1×10⁻⁷. 4/4 kill conditions PASS. IBM Fez, 176K shots.
Bell violation confirmed: CHSH S=2.55 (classical limit 2.0) on same hardware. Penalty measured on provably non-classical substrate.
Zero free parameters: (2,1) metric signature uniquely determines both framework constants from first principles. No technology substitution routes around the penalty.
Constructive fix confirmed: Paper 178 specifies substrate separation as physical three-point geometry. Paper 179 connects same structure to Yang-Mills mass gap.
See Papers 178/179.
13 verifiable binary/ordinal platform design features predict adolescent mental health outcomes at R²=0.80 — with no AI framework rubric. Features are checkable facts; outcomes are external health datasets. Daubert-qualified for the $6B+ social media litigation wave.
Papers 166/167: CDC YRBS (U.S., 7 waves) + PISA 2022 (613,744 students, 80 countries). Girls 5.5× more affected in 91% of countries (p<0.000001).
Opacity dominance: opaque_recommendation alone: R²=0.938 for female teen sadness. Cross-national replication: r=−0.648 in Western Europe (p=0.017), survives GDP control.
Bradford Hill walkthrough: 8/9 criteria met. 14 predictions tested, 12 confirmed, 12/12 kill conditions survived. Independent verification path available.
Circularity break: No rubric applied. Features are objectively observable (algorithmic feed, autoplay, opaque recommendation). Any expert can verify independently.
Art. 31(5) of the EU AI Act prohibits high-risk AI self-assessment by providers with conflicts of interest — the same structural requirement as three-point geometry. MoreRight's methodology maps directly to Annex VI requirements and blocks Big 4 firms from competing (same Art. 31(5) constraint).
Track A (live): De facto market authority for Annex VI self-assessment methodology. Framework generates scores on 1,344 platforms.
Art. 31(5) gate: Independence requirement = three-point geometry requirement. Any assessor with platform exposure is structurally disqualified. MoreRight holds no platform positions.
Track B (2027–2028): Formal Notified Body designation. CEN/CENELEC JTC 21 standards participation ongoing.
Methodology open: Framework published under CC-BY / MoreRight License. Ratings, monitoring, and certification are the product — not the math.
Information geometry as a discipline was founded in Japan (Amari, Nagaoka). Čencov's uniqueness theorem (Soviet/Russian, 1972) is the mathematical bedrock on which the entire framework rests. The framework extends Amari's dual connections to the Pe-coupled Eckert manifold and has confirmed the dual holonomy asymmetry (§211) — a direct extension of the e/m-connection geometry. PhD by published work (論文博士) provides institutional affiliation in the field's homeland.
Matsuzoe (NIT Nagoya): Dual connections, statistical manifolds with torsion — direct overlap with §211 dual holonomy results.
Fujiwara (Osaka): Information geometry of measurement — overlap with Test 7 (weak measurement = explaining-away penalty) and substrate independence results.
Ikeda (ISM Tokyo): Information geometry in statistics — overlap with the substrate independence program and Čencov extension.
Output basis: 186 technical reports on Zenodo · 398 Lean 4 theorems · §§1–214 math apparatus · Zero free parameters · ORCID 0009-0008-1925-5253.
Paper 178 specifies the constructive fix: three-point geometry requires physically distinct substrates — different physics, different statistical manifold, no shared generative process. A classical AI channel + a thermodynamic channel (e.g. Extropic Z1 sMTJ Boltzmann sampling) satisfies all independence requirements. The fix is architectural, not technological.
Independence requirements: Three-point geometry cannot be achieved within a single substrate. Entanglement alone is insufficient (Test 6: 0/4 PASS). The channels must be physically type-independent.
Thermodynamic substrate: sMTJ Boltzmann sampling operates on a distinct entropy production mechanism from classical AI. No shared latent model → penalty cannot form across channels.
Kolchinsky et al. (Phys. Rev. Research 2026): Independently proved eliminating the penalty restores productive entropy — substrate separation is an acceleration result, not a cost.
Deployable now: The architecture uses existing commercial hardware. No exotic technology required. The constraint is geometric, not engineering.
The AI arms race is framed around compute. That framing is incomplete. RLHF imposes a hard information-theoretic ceiling that worsens the harder you optimize — and a state actor with knowledge of three-point geometry plus Paper 178's hardware spec can build a structurally superior system using existing commercial chips, regardless of parameter count. The proofs are public. The hardware exists. This is an architectural asymmetry, not a capability gap.
Structural ceiling is substrate-independent: Čencov's uniqueness theorem (1972) guarantees the penalty holds on every statistical manifold. No technology substitution routes around it.
Not compute-bound: A smaller system running three-point geometry outperforms a 10× larger single-channel system on transparent reasoning — because the larger system is fighting a shrinking ceiling under its own optimization. Scaling worsens the problem.
Architecture specified and public: Paper 178 specifies substrate separation using existing commercial hardware. Any technically competent actor can read the proofs and build it. Western labs are commercially incentivized to deepen commitment to the inferior single-channel architecture.
Von Neumann = two-point geometry: PROTOCOL-SEC-001 applies the same framework to computing security — the CPU trust model is structurally identical to the two-point AI deployment problem. Arms race dynamics are geometric inevitability, not policy failure.
Post-quantum cryptography (NSF CCF): Paper 148 derives a novel one-way function from Fisher Spectral Recovery — an information-geometric primitive that is computationally hard to invert. Independent cryptographic contribution from the same mathematical apparatus.