GeWu.ai
An observatory for AI education design.
AI education design is scattered and hard to learn from. Some sits inside repositories — prompts, guardrails, architecture choices — readable only by reconstructing each system by hand. Much never touches a repository at all: competency frameworks, institutional policy, curricula, assessment redesign, and the research that should inform all four.
This observatory brings that material into one library, organises the decisions inside it, and puts them to work on a design you are actually making — three kinds of work, which is also how the library is arranged: building agentic AI for teaching; teaching AI as subject matter; and redesigning assessment so it still evidences learning now that models exist.
Two pieces of my own underpin the work. The Simulators treat a language model as an object of empirical study rather than a tool, following Probing the Stochastic Machine; the evidence checks throughout follow Can AI be Trusted? Beyond them the library draws on published learning-science research and open benchmarks, cited in place throughout the Knowledge Graph.
— Tian Zheng, lab lead, aiX Lab
Read this before you rely on it
It changes. Pipelines are re-run and corpora re-curated, so a link you saved may have been reorganised underneath it.
Much of it was written by a model, not a person — plans, critiques, summaries, extracted elements. Models make mistakes fluently: a fabricated citation reads exactly like a real one. Nothing is checked by a person unless it says so.
Review anything you intend to teach from or cite. Evidence tiers and amber flags say what the system knows about its own grounding; they are not a substitute for reading the original.
Where the name comes from
格物致知
gé wù zhì zhī
From the Great Learning (大學), and a central problem of Song Neo-Confucian thought: knowledge is extended by close investigation of particulars, not by reasoning from first principles about them. Zhu Xi read it as examining the pattern in individual things until the underlying principle becomes clear. In nineteenth-century China the shortened form 格致 denoted natural science, before 科學 displaced it.
The phrase divides in two. So does the system.
格物
gé wù — to investigate things
GEWU, the analyzer
Reads an existing project — repository, prompts, configuration, documentation — and extracts a profile, its design blocks, the decisions inside them, and the risks they carry. It never executes what it reads. It also reads in the other direction: Z² calls GeWu back to critique a design against the library — where the theory does not support it, which patterns it is missing, and where the evidence is thinner than the plan assumes.
致知
zhì zhī — to extend knowledge
Z², the design co-pilot
Reasons over that library toward something new: which design blocks apply, what the tradeoffs are, what will hold up in the setting you deploy into. The name is the two Z-sounds of zhì zhī — Z, squared.
The mark
The structure of diamond: valuable because it is strong, and strong because of how it is assembled. What rigorous knowledge should look like.
The box is GEWU — eight points, twelve edges, a cubic cell drawn flat so every side is visible at once. The scope of what is examined, and what holds the knowledge.
The bonds are Z² — four reaching from the centre to four corners, the tetrahedral bonding that gives diamond its hardness, and connecting the dots. Nothing is added to the cube. Its strength is a matter of which points are joined.
How to use it
There is no required order, but reading before designing tends to work better than the reverse. Each tab answers a different question.
A design move rarely appears once. Clusters and themes group the recurrences, linked to learning-science theory, published findings, and the settings a design can run in. Search by meaning rather than wording, and save a search worth repeating as a tag.
Cases decomposed into blocks, then into individual elements: the sentences of a prompt, the guardrails, the tool bindings. How a system was actually assembled, rather than how it was described afterwards.
GEWU reads a repository — never executes it — and extracts a profile, its design blocks and the risks it can identify. Material with no repository to read comes in through a YAML template instead. Adding happens on the machine that builds the library, so the tab is closed elsewhere.
Restricted to elements whose effect must be seen to be credited: refusals, gates, withholding, character consistency. Elements that read identically with the behaviour on or off stay in the Case Library to be read.
Three tabs, each labelled Teach with AI — a discipline is the subject, AI the instrument — or Teach about AI, where the model itself is. Agent conversation forks a teacher–student exchange on one guardrail. Token probabilities reads the model's distribution over your answer set; Branching walk traces probabilities into words and compares five decoding strategies.
A consultation covering subject, learners, setting and deliverable, ending in a design plan and a paste-ready artifact. The plan is a draft, not a verdict: edit it directly, or tell Z² what to change and get it revised. Sessions run live with Tian Zheng; past ones are open to read, and plan and transcript download as Markdown.
Who it is for
Faculty, instructional designers and researchers thinking about how to build, or evaluate, AI systems for teaching. It is a site for consultation and demonstration.
It is not a student-facing tool. There is nothing here for a class to use directly, and it is not intended to be put in front of students.
Getting an account
Reading is open. Sign in and the cases, the knowledge graph and any shared consultation are yours to browse. Running the parts that call a commercial model is the level above, and is asked for — say what you teach or are building and we will get back to you. Or supply your own API key in Settings, which opens the Simulator immediately.
Request access →Accounts carry one of three roles, because the parts of this site that call commercial models are billed to the project:
- Reader — browse the library and the Knowledge Graph, read your own consultations, and read any that their author has agreed to share.
- User — also run Elements Lab runs and Simulator probes, billed to you if you supply your own keys in Settings. The Z² Workshop runs as a live session, so starting one needs the administrator role or the word issued for a session in progress.
- Administrator — also run consultations and the pipelines, read every consultation, add to the library, and delete material.
Holding an account means accepting the Terms of Use, which set out what may and may not be done with the material here. The library summarises other people's repositories and papers: it is available to read, not to republish.
What happens to what you type
Consultation sessions are saved. Transcripts and plans are retained so the system can be evaluated — whether the consultation actually helps is the research question this project exists to answer.
Simulator sessions are not retained. Probes and decoding runs are never written down; agent conversations are held only while they run and deleted within 24 hours.
Prompts reach Anthropic and OpenAI to be processed, so your exposure is the same as using ChatGPT or Claude directly — or put your own key in Settings and have it run under your account. Full detail in the privacy note.
What it does not do
- Execute analyzed code. Files are read and classified, never run.
- Decide for you. Extracted blocks record another team's choice, not a recommendation.
- Judge fit. Whether a design move suits your students is outside what the library can determine.
- Overstate testability. Decisions that leave no trace in a transcript are marked read-only.
- Deploy. It produces designs and artifacts to be built elsewhere.
Acknowledgements
The library is built from public repositories, frameworks, policies, curricula and research released by the teams and scholars who made them. Whatever design judgement it contains is theirs; this system only organises it, and cites it in place throughout the Knowledge Graph.
Evidence-law examples in the Simulators and Elements Lab come from LegalBench, an open benchmark written by legal professionals.
Built with Claude and OpenAI models, Next.js, FastAPI and PostgreSQL with pgvector. Development was itself an exercise in the kind of human–AI collaboration the system is about.
Run on the lab's own API accounts, without external funding — which is why reading is open but running is asked for, and why supplying your own key sidesteps the constraint entirely.