AI-Ready Knowledge Infrastructure
Whether AI works well in an organization turns on more than the model: it turns just as much on how well the documents and data the model consults have been put in order. I use the term AI-Ready for the policy of shaping documents, data, and algorithms so that an AI system can reference and reuse them, and this topic is about designing that. The experimental evidence comes from Which Algorithm Specification Formats Help Language Models Implement Machine Learning Algorithms? (Kato and Kato, 2026, arXiv:2607.03158).
Problems addressed
I think of an AI-Ready knowledge base as having two layers. The first manages the source documents, their metadata, where they came from, and the route by which they get updated. For an AI system to show the grounds for an answer, someone has to be able to trace which document and which version it drew on. The second layer holds algorithm specifications that state the inputs and outputs, the computational steps, how numbers are handled, and the boundary conditions. Papers and documentation usually leave the implementation choices implicit, and that is where an AI system trips.
To test the second layer, we wrote the same algorithm in seven formats (prose, two kinds of pseudocode, Markdown, YAML-like, JSON-like, and Python skeleton code) and then compared how accurately language models implemented each version. Across three models and five machine learning tasks we generated 4,020 implementations. Where information was missing, the format mattered a good deal. Where the specification already carried what was needed, the gap between formats closed as the model grew more capable. What decides the outcome is whether the specification states the choices an implementation has to make; how it looks counts for much less.
Building this out takes more than installing RAG, and no single piece of research makes a whole organization AI-Ready. What this study offers is experimental evidence to design against.
Core research
The Algorithm Specification Formats page describes how the experiment was designed, what it found, and the writing guidance that follows.
Research thread
- 2026
- AI Economist Agent. A worked example of the design, holding economic data and theory as a knowledge graph that the agent consults to assemble its analysis.
- 2026
- Knowledge Infrastructure for the AI Era: OKF, GBrain, and Others (in Japanese), a Zenn article for practitioners on how to record the knowledge an AI agent consults and how to put it to use.
Relation to earlier work
Evaluation of code generation by language models has grown up around implementing a function from its specification, following the study of Chen and coauthors (2021, Evaluating Large Language Models Trained on Code). Sclar and coauthors showed that small differences in the surface form of a prompt can move performance a long way (2023, Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting). Our study complements both. It takes specifications for machine learning algorithms as its object, and it measures implementation accuracy by asking which information the specification states rather than how the specification is laid out.
Where it applies
If you want an AI system to read your internal documents and data, or want procedures and model specifications recorded in a form it will not misread, this work gives guidance on how to write them and how to manage them. For inquiries about designing a knowledge base or getting documentation into shape, write to mkato-csecon@g.ecc.u-tokyo.ac.jp.