No single model saves us
Entity resolution, graph neural networks, classical solvers, language models, each used where it is strongest and coordinated rather than collapsed into one god-model.
CTI
CRITICAL THINKING
INFRASTRUCTURE
INFRASTRUCTURE
Infrastructure for asking good questions quickly.
The bottleneck in high-stakes decisions is rarely rigour. It is speed. The better method loses to the spreadsheet because the spreadsheet answers today. CTI is an attempt to make the rigorous method answer in minutes, and an honest account of what is still open.
Research · work in progressA research direction from Second Order AI Ltd.
The idea
A superior method that arrives too late is not superior.
Under operational pressure, teams drop the rigorous method for intuition and a spreadsheet. The method is sound. It just answers in weeks, and the decision is due now. The rigour was never the problem. The wait was.
CTI asks what it would take to keep the rigour and lose the wait: a defensible answer in minutes, in the language the decision-maker actually uses.
The sketch has four layers: a spatial knowledge graph, a domain graph, a solver layer, and small purpose-built tools. A language model coordinates the specialised models, rather than one model doing everything.
Above the technical stack sits a social layer: a model of how decision-makers frame the problem in the first place. Frame it wrong, and the best solver still answers the wrong question.
The social layer
Model the decision-maker, not only the network. How a problem is framed decides which answer counts as good before any solver runs.
Written as testable claims
The series keeps a hypothesis register: what is claimed, what is tested in commercial work, and what remains an open question.
The series · seven parts
Read Critical Thinking Infrastructure
The full argument lives on Elias’ site, written up as a seven-part series. It draws on decision research (Tversky and Kahneman, naturalistic decision-making, the no-free-lunch theorems) and maps the idea onto real public-good and venture cases.
“The right answer to the wrong question is still wrong.”