
Issue 23 · Shared Meaning
If this landed in spam or Promotions, drag it to your inbox and hit reply with one word. It tells your inbox this is mail you want, and it helps every issue after this one reach you.
Hi friend,
Last issue ended with knowledge graphs, ontologies and a promise to show you the proper layout of CHARM. This issue keeps both promises, and it starts with a small problem that causes a large share of agent failures: two parts of a system using the same word to mean different things.
So this week covers what a knowledge graph and an ontology are for, the five layers of CHARM, how I picked this week's news from my own knowledge graph, and a book I want to read with you.
This week
🔨 Built Shared Meaning, what knowledge graphs and ontologies do, the five layers of CHARM, and how my news graph picked this week's stories
⚡ One Slot, the code review and cleanup slot, Gemini 4 Carbon against Claude Fable 5.2, and my verdict
🌎 In the News, four quick stories, Nvidia's CEO change and open source, OpenAI's Control Room and ChatGPT Wallet, and DeepMind's AlphaProof Code
✅ Before You Ship, name your ten words
💻 Prompt for Productivity, build a starter vocabulary
🗓 Open Calls, Claude Loves Lovable, the Orinyx MCP 0.1.5, and four picks, Global AI Kansas City, Stanford's virtual AI+HEALTH, a graph hackathon and a free founder sprint
🔭 Next From Me, how to build your own knowledge graph to keep up with the news
A knowledge graph stores things and the relationships between them, so a machine can follow the links instead of guessing from text. An ontology is the agreement underneath it, a written definition of what kinds of things exist in your world and how they are allowed to relate. The graph holds the facts and the ontology holds the meaning.
I keep reaching for them because an agent is fluent before it is precise. Ask two agents about an "active" customer, and one counts anyone who logged in this year while the other counts anyone with a paid plan, and both answer with confidence. Nothing is broken in either one. They never agreed on the word.
I built a small knowledge graph to keep up with the news for this newsletter, because I am working on a podcast version of Artificially Designed, and this week's news came out of it. This morning I refreshed it from the Google DeepMind, OpenAI and Rundown AI feeds, and it now holds 1,509 articles linked to their sources and topics by 3,657 relationships. It makes two kinds of claim, that an article is about a topic and that a source published it, and nothing stronger. It does not conclude that a company launched a model because a headline says so. All 1,509 story candidates start at Needs Review, and I decide what moves forward. The graph narrows the pile and I do the judging, and the same split is why a person owns the last decision on anything I build.
That gets me to CHARM, my framework for designing human-centered, governed agentic systems. The name stands for Human-Centered Architecture for Responsible Multi-Agent Systems, and it is built on five layers.

Connect. Integrate people, systems, data, tools and workflows.
Harmonize. Establish shared meaning through semantic models, ontologies and knowledge graphs. This is the layer this issue is about.
Augment. Design intelligent agents that support human capabilities and coordinate work.
Regulate. Embed governance, permissions, boundaries, accountability and human oversight.
Monitor. Continuously evaluate performance, reliability, safety and human outcomes.
These are interconnected capabilities and not five steps in a row. Human-centered design and governance should shape the system from the first discovery conversation to deployment.
CHARM brings together four disciplines that usually get worked on separately: human-centered design, semantic architecture, agentic systems engineering, and continuous governance and evaluation. The contribution I am making is putting them into one design method, and whether that method earns its place has to be shown through real implementations. I am not asking you to take it on faith.
Under all five layers sits one constraint. Human agency comes first. Before anything gets automated, the questions are whether it deserves automating, who keeps the decision, how a person steps in, and whether the people using it end up better off. The principle I work from is three short lines.
Design for human agency. Architect for shared meaning. Govern for accountability.
CHARM applies across industries. It fits enterprise AI, where it shapes internal agents, organizational knowledge and workflows, and healthcare, where it shapes clinical information systems and regulated AI. It also fits engineering work such as construction, environmental and operational systems, education, AI-assisted learning and workforce development, AI assurance, which covers agent verification, oversight and auditing, and multi-agent systems, where it shapes coordination, semantic exchange and orchestration. Orinyx is its own product and applies some of these principles in clinical settings, and CHARM belongs to charmthirteen.
One more thing before the news. The Harmonize layer has a book that covers the same ground, Ontology Pipeline: A Framework for Knowledge Engineering by Jessica Talisman, published by Technics Publications. It walks from controlled vocabularies through taxonomies, thesauri and ontologies to knowledge graphs, and it comes from the way librarians have organized meaning for a long time. Technics Publications' page for the book
⚡ One Slot
The One Slot Rule is that every job gets one slot, and a new tool has to take it from whatever holds it now. This week the slot is code review and cleanup, and the tool asking for it is Google's new Gemini 4 Carbon.
What is the job. Reading code that other tools wrote, finding what is wrong with it, and cleaning it up before I push. Claude Fable 5.2 holds that slot for me today, and I am still very much using it. For overall planning I use smaller models like Claude Opus, which keeps the expensive model for the job where it earns its cost. That refines last issue's line that Claude plans. The planning now goes to Opus, and Fable 5.2 sits beside me for the review.
Can it do that job, not can it do more. Google released Carbon on November 10 and says it is built for multi-step coding and debugging, with an introductory price of $2 per million input tokens and $10 per million output tokens. Those are Google's claims about its own model, so I looked at an outside comparison. On BenchLM's leaderboard, updated November 15, Fable 5.2 is first at 93.84, Carbon is second at 89.56, and Claude Opus 5.5 is third at 86.81. For coding specifically, BenchLM scores Fable 5.2 at 89, Carbon at 84 and Opus 5.5 at 83. BenchLM prices Fable 5.2 at $8 and $40 per million tokens, so Carbon keeps about 95% of the top score at a quarter of the output price. BenchLM also labels Fable 5.2's rank Estimated, with a wide uncertainty range that overlaps Carbon's, and says the ranges do not establish how confident any rank is. So there is a lot still to work out, and I am not moving off Fable 5.2 on a leaderboard.
What does it touch that is hard to undo. Carbon is also the model behind a preview feature in Google Antigravity called /swarm, which turns one prompt into a coordinated team of agents working in parallel, isolated workspaces, with adversarial checkers testing each change. What is hard to undo is a change landing in a real project, which is why the review slot exists. I use a related idea in my own work. I tried Claude Code's ultracode, which spreads a job across many agents, and it used up my usage and made me reset it twice. I would do it again, and I also learned how quickly a multi-agent run spends tokens.
What would have to be true to let it run unwatched. Same two conditions as last issue, an approval gate on every action that cannot be undone and a person reviewing before anything merges. A team of agents does not remove the second one. It makes the review more important, because there is more to read.
My verdict is that Fable 5.2 keeps the slot, Opus keeps planning, and Carbon stays on trial. I will tell you what I find.
🌎 In the News
Four stories from the last week, kept short on purpose. The deep dive is coming, and I will tell you where soon.
🟩 Jensen Huang steps down at Nvidia. He resigned as CEO on November 12, and CFO Colette Kress is interim CEO while the board searches. I am watching two things, Nvidia's earnings on Tuesday, the first test under new leadership, and open-source AI, where Huang was one of the loudest supporters of Nvidia's Nemotron open models. People disagree on whether that support survives him, and my read is that if you build on open models, make sure you can swap one out. The story is still moving. Nvidia's announcement and TechCrunch's open-source analysis
🛑 OpenAI shows how it stops its own models mid-run. Control Room tracks its research workloads and can halt a job's output within 35 seconds, and if it goes down, everything stops. I read it as a design rule, put the off switch where the code changes least. This is OpenAI describing its own system. OpenAI's Control Room post
💳 ChatGPT agents can now pay, with you approving every purchase. ChatGPT Wallet, built with American Express, Mastercard and Visa, has you approve each purchase with a security check, and the agent never sees your real card number. It is the next step after Claude's read-only finance feature, the person approves each irreversible action. OpenAI's ChatGPT Wallet announcement
🔐 DeepMind used AI agents to prove the code behind Signal's encryption is correct. People wrote and reviewed the specs, agents wrote the proofs, and a proof checker verified them. I like that the written spec is shared meaning a machine can check against, and people stay responsible for what it says. This is DeepMind describing its own work. DeepMind's AlphaProof Code post
✅ Before You Ship
Name your ten words. Before you point an agent at your own data or documents, write down the ten words your team uses most, such as customer, active, order, resolved or approved, and give each one a one-sentence definition.
Then ask two people on your team to define each word without looking at yours. Where the definitions differ, you have found the place an agent will quietly go wrong. Fix the definition first, then build.
💻 Prompt for Productivity
Build a starter vocabulary from your own documents.
Act as a knowledge engineer who builds controlled
vocabularies. I will paste text from my own documents. Build
a starter vocabulary from the terms that matter to my work.
For every term, give me:
1. Preferred label: the one name we will use
2. Definition: one sentence that says what it is and what it
is not
3. Alternative labels: other names or abbreviations people use
for the same thing
4. Parent term: the broader category it belongs to, using an
"is a" test
5. Source: a short quote from my text where the term appears
Rules:
- Use only my text. If a term is not in it, do not add it.
- If two terms look like the same thing, or one term is used
with two meanings, list it under "Needs a human decision" and
say why. Do not pick for me.
- Mark every definition you inferred instead of found as
"inferred".
Text: [paste, with names, account numbers and anything
confidential replaced by <REDACTED>]WHEN TO USE IT. Before you build an agent, a knowledge base or a prompt library on top of your own documents, and again when a new team or system joins.
WHAT IT DOES. It turns scattered wording into one list with definitions, and it hands you the conflicts instead of resolving them for you.
TIP. Read the "Needs a human decision" list first, because that list is the real work. Then ask someone from another team whether they would define the top five terms the same way.
🗓 Open Calls
From me
Wednesday, November 18, 8:30 to 10am Central. Claude Loves Lovable, my hands-on workshop series with Central Exchange, and this week's session, Build Your Website with Claude + Lovable, is the last one in the series. Registering covers the sessions, so please register only if you plan to attend live. It is virtual, no coding or prior Claude experience is needed, and it is free for Central Exchange members and $25 for non-members. Register through Central Exchange
The Orinyx MCP, version 0.1.5, out November 15. Orinyx checks the medication orders and claims your agent produces against FDA label text, cites the section it checked, and says plainly when the label gives it nothing to check. This release is the first one on PyPI, so you can install it with pip install orinyx-mcp or run it with uvx orinyx-mcp stdio. It adds 35 medications, covering high-alert hospital drugs, chemotherapy, and heart, blood pressure and metabolic drugs, each mapped to its FDA label and still awaiting physician review. It also hardens sign-in and credential handling. It is a limited testing release on synthetic data and is not for clinical use. Read the 0.1.5 release notes or join the Orinyx MCP waitlist
Four picks for the week
Thursday, November 19, 6 to 8pm Central. Global AI Kansas City, the chapter's November meetup, in person in Kansas City. Event page
December 8 and 9, 8am to 2:30pm Pacific. Stanford AI+HEALTH 2026, a virtual conference on AI in health and medicine, with regular registration at $350 for academic, healthcare, nonprofit and government attendees, $550 for corporate and $50 for students. Register
Through December 11. The Relate by Neo4j Hackathon, an online, month-long build around the theme "Context, Connected," where you show how connected data helps agents and apps find the relationships that matter. Relate is an open-source embedded graph database for AI agents. The Devpost page lists a December 11 deadline at 6pm Pacific. Devpost page
Apply by December 1. INSEAD AI Venture Lab's AI Founder Sprint 2027, a free, virtual, 10-week program for early-stage founders from idea to pre-Series A, with no equity taken and non-dilutive prize money at Demo Day. It runs January 27 to April 2, 2027. Program site
Tool verdicts, templates and reference material live on the site at charmthirteen.com.
🔭 Next From Me
Next time I will show you the knowledge graph I built, a simple project, and how you can build your own to keep up with the news.
💌 One Ask
I am starting a book club to read Ontology Pipeline by Jessica Talisman, and I would love for you to read it with me. If you want in, hit reply with one word, and I will send you the details as soon as I have them.
Lex
