The 16% Solar Variance: When Green Lights Bleed Balance Sheets
Sibaya Casino generated 2.96M kWh expecting a 20% grid bill offset, but usage only dropped by 5.19%. Asoba's physical audit disproved the "weather excuse," proving 92.6% of lost solar yield was silent hardware underperformance costing R1.15M ($69.3K) annually.
Bounded Intelligence: Coupling a JEPA World Model With a Retrieval-Grounded Reasoner
An anomaly score confirms behavior has departed from normal — it doesn’t say why, or what to do. Closing that gap safely means refusing to let one model do both jobs.
The Utility Ledger: A New Category for Energy Infrastructure
SCADA monitors but doesn't settle. ERP manages finances but doesn't understand physics. No single system connects technical performance to commercial outcomes—from feasibility analysis through carbon credit issuance. The gap is not a feature request. It is a category.
The Energy Map Is the Power Map: From Bronze Age Tin Routes to Petrodollar Collapse
The shape of global power has always been determined by who controls the trade routes for the era-defining energy resource. The Bronze Age collapsed when its trade network became too fragile. The petrodollar system is the modern equivalent.
The Uranium Corridor: French Militarism, Sahel Resources, and the Next Theatre of Great Power Competition
The Iran War is the world’s primary energy conflict theatre. But the Strait of Hormuz closure is triggering something larger: a scramble by every major non-US power to secure energy resources outside the oil-and-gas architecture.
The Iran War is the single largest accelerant of the global renewable transition. But the Israeli right wing’s actual strategic objective — an oil pipeline through Israel to Europe — requires a forever war to deliver a commodity the intended customer is racing to stop buying.
System 1 at the Grid Edge: A Controlled Comparison of Per-Inverter JEPA and Frozen LLM-Backbone World Models for Energy Telemetry Anomaly Detection
Controlled head-to-head on 1.1M SCADA records across seven Sibaya inverters: per-inverter JEPA beats the frozen LLM backbone by +0.15 AUROC and 2.7x average precision, training in 3-4 minutes per asset on edge hardware
October 2026
Per-Inverter World Models for Unsupervised Fault Detection in Distributed Solar Fleets: A Joint-Embedding Predictive Architecture Case Study at the Sibaya Solar Site
JEPA-based per-inverter world models across seven Sibaya inverters (18-month continuous 5-minute SCADA telemetry, 1.1M records), achieving mean 0.8906 AUROC, 78% false alarm reduction, and +0.76 AUROC margin over physical persistence
September 2026
Epistemic Fine-Tuning of Open-Weight LLMs for Deep Research: Nehanda v3.1 and the FACTS Grounding Benchmark
Targeted fine-tuning of a smaller open-weight model (Nehanda v3.1) for grounded RAG synthesis, scoring 82.21% on DeepMind's FACTS Grounding benchmark
May 2026
RAG Synthesis on a Native Multimodal Architecture
Stacked fine-tuning methodology for Nehanda expanded to qwen3.6 VL architecture
March 2026
Generative AI LLMs vs. Specialized Neural Networks for Decision-Making and Probabilistic Inference in Distributed Energy Systems
Where large language models vs specialized neural networks serve microgrids, VPPs, and grid management
March 2026
Epistemic Robustness via Stacked Cognitive Training: Improving LLM Grounding for Policy and Intelligence Analysis
Stacked fine-tuning methodology for epistemic honesty in high-stakes policy and intelligence domains
January 2026
Energy-at-Risk Assessment of a 2.31 MW Commercial Solar Installation
88-day AI-driven anomaly detection and loss quantification across a rooftop PV system in Durban
October 2025
The Intelligence Layer: A Technical Framework for Predictive and Decision AI in Distributed Energy Asset Management
AI framework integrating predictive forecasting with automated decision execution for renewable energy
Asoba