SEPT Principles — Overview

Why SEPT?

SEPT emerged from years of hands-on work validating, correcting, integrating, and migrating data. In long-running systems, repeated patches and migrations can leave records whose original meaning or change history is difficult to recover. Integrating data from different sources can also obscure differences in definitions and classifications.

Generative AI can produce answers from large amounts of information. But storing every LLM response does not automatically produce a reliable knowledge system.

What matters is selecting and relating useful information, abstracting it into knowledge that people can understand and verify, and keeping its connection to the source data. AI should be able to use that knowledge too.

A rule inferred from several examples may later encounter an exception. We may need to revise its assumptions, add conditions, and verify it again.

Knowledge should not be treated as a finished product. It should remain open to new questions and evidence.

What is SEPT?

SEPT consists of four design principles to consider when creating, abstracting, generalizing, integrating, updating, sharing, and reusing knowledge, so that its meaning and supporting evidence are not lost. The same principles can be used to evaluate knowledge and knowledge systems.

Plural — preserve meaningful differences

Plural is especially important to my approach.

While working with Japanese COVID-19 open data, I found that prefectures used different age categories. “Under 10” and “not yet enrolled in school” do not mean the same thing. Likewise, “90 and over” contains less detail than separate “90–99” and “100 and over” categories.

Mapping these values to a single category without recording the original meaning may lose information. Fine-grained categories can be aggregated when their definitions are compatible, but a broad category cannot be split into precise subgroups without additional data.

Differences also arise in interpretation. Historians may reach different conclusions from the same sources, and business teams may make different decisions using the same market evidence. We do not always need to force these perspectives into one answer. This does not mean treating all claims as equally correct.

Plural means retaining meaningful differences so they can be compared and evaluated.

Use SEPT while creating knowledge

SEPT is not only a checklist for a finished system. It informs decisions from the moment knowledge is created: extracting information from documents, generalizing observations, integrating data, or recording decisions.

We do not need to build a complete knowledge model before starting. Begin with the knowledge needed for the current question, use it, validate it, and extend it when new questions or evidence arise.

Question → Minimum Knowledge → Apply & Validate → Identify Gaps → Extend & Generalize

This is one practical, iterative approach—not a procedure required by SEPT. And Minimum Knowledge does not mean Minimum Data: retain access to sources and evidence as appropriate.

Technology-independent principles

SEPT is not tied to a particular AI system or database. Graph data models are one useful way to connect claims, sources, interpretations, changes, and decisions.

Approaches such as Knowledge Versioning and Knowledge Graph Diff may help implement these principles, but neither is a required part of SEPT. Knowledge Versioning draws on ideas from Git without implying use of Git itself.

The goal

SEPT does not mean saving everything or designing a complete knowledge system upfront.

Start with the knowledge you need. Preserve its meaning, different interpretations, changes, and evidence, so people can understand and verify it, AI can use it, and the knowledge can continue to evolve.

For detailed definitions, scope, and evaluation criteria, read the SEPT Basic Definition.