Artificially Designed · Issue 18: The Alias

Issue 28 · The Alias

An agent can cite its source perfectly and still be wrong about what the word meant.

Hi friend,

Last week I asked whether you could explain why an agent did the thing that worked. Keep your answer close, because this issue is about the part that never shows up in a log.

Stack Overflow's 2026 Developer Survey, out October 6, maps where developers let AI in. Most use it to write code in areas they already know (67%) and to debug (61%), but only 20% let it near production, meaning deploying, operating or troubleshooting live systems. And 93% say that trusting what AI hands them requires source attribution.

Attribution is the right instinct, and it has a hole in it. A citation tells you where a word came from, and it can't tell you whether the agent meant what you meant by it. I spent last week inside that hole. It is the first of the three layers I promised you, Design for Meaning.

This week

The build I see most treats a name as a key. A customer types "Acme," the database holds "ACME Inc.," a fuzzy match joins them, and in most products a near miss costs you a duplicate record.

In medication a near miss costs you a different drug.

Last week's issue told you about a public benchmark of about 65,000 free-text medication orders, the same drug written as a brand, a generic or a pharmacy shorthand, where our checker understood only one of the three. The obvious fix is smarter matching. Map every brand to its generic, let a fuzzy match catch misspellings, and watch the decline rate fall. RxNorm, the National Library of Medicine's drug vocabulary, will hand you the brand names for every drug Orinyx covers. We had that list, 913 names, and nobody had reviewed it one by one.

Here is what testing taught us about loose matching, in three names.

Penicillin. An early version matched the bare word by its first letters and landed on oral penicillin V potassium. A correct order for an injection would then have been flagged for the wrong route, with full confidence.

Ferrous gluconate. The matcher stripped the salt, kept "ferrous," and landed on ferrous sulfate, a different salt with a different dose basis. Every dose check after that would have used the wrong label. That was a guess dressed up as a lookup.

Metoprolol. One word, two products, two labels, succinate extended-release and tartrate immediate-release. Resolve it to either and you have decided something the order never said.

None of these is a model hallucinating. Each is a mapping that looked reasonable and quietly picked a meaning. Brand names do it at scale, because one brand can cover a plain tablet, an extended-release tablet and an injector pen, and "brand equals generic" throws the formulation away. Out comes a confident wrong answer, with nothing in the output to say so.

The word for what was missing is ontology, and you can ignore the word. For a founder it means the agreed meaning of each word your product depends on, written down before your data, your interface or your agent uses it. Your product already has one. It lives in column names, dropdown labels and the prompt someone wrote at midnight, and when those disagree, your agent picks a meaning for you.

This is also where health AI companies die, and it is rarely the model. Cydoc's founder shut the company down after seven years and wrote the most useful postmortem I have read this year. Cydoc built screens for history, medications and allergies that clinicians never wanted, and the diagnosis is blunt: "We hadn't properly thought through the clinical workflow." Read it

Three moves hold, and none of them needs a drug vocabulary.

Write the definition before the field. One sentence per core noun, agreed by whoever builds and whoever sells, kept where your prompts can read it. If "customer" means an organization with a paid plan, the column, the button and the system prompt all say so.

Let a person sign every new synonym. A synonym is a decision. An alias list grows because someone read each proposal and wrote down why it holds, never because a match score crossed a line. RxNorm's own approximate-match tool returns a score and leaves the cutoff to you, which is honest, because the cutoff is a decision about meaning.

Make "I don't know" a valid answer. When a name is not on the list, the system stops and says it cannot check this as written. That answer looks like failure in a demo, and in production it is the only one you can audit.

On October 10 I made that a written rule for Orinyx. No fuzzy matching, and no live vocabulary lookup when a request comes in. Recognition grows only through a review batch, a worksheet with one row per proposed brand name, the products it actually covers, and a flag on anything that changes the formulation. An accepted row has to carry the reviewer's own note on the formulation, written by a person and never generated. Every row carries the fingerprint of the file it was cut from, and a worksheet that breaks any of that is refused whole.

A brand that has left the market, the kind that still turns up in old charts, now gets an honest answer too. Orinyx recognizes it as a real drug it does not cover and abstains instead of guessing.

It is slower, and the decline rate falls a row at a time. But every name Orinyx recognizes now has a person's sentence behind it, which is the record of meaning I went looking for last week and couldn't find.

The One Slot Rule is that every job gets one slot, and a new tool has to take it from whatever holds it now. The tool is mine again this week, so hold me to the four questions.

It is the Orinyx MCP, a connector that lets an AI agent check a medication claim against the cited FDA label text before it acts. The rules are deterministic, and no language model takes part in a verdict. It is free, and it is in design sessions with physicians right now, so nothing below is a claim about clinical accuracy.

What is the job. Before an agent writes or suggests anything about a medication, check each claim against the label it cites. In most builds that slot is held by nothing, or by the same model checking its own work.

Can it do that job, not can it do more. It answers one claim at a time, supported, contradicted, unreadable or abstain, and a supported or contradicted answer names the label section it was checked against. Supported means the claim agrees with the label text. It never means an order is right for a patient, and it certifies nothing as safe.

What does it touch that is hard to undo. Nothing. It reads labels and returns answers. It writes to no chart, sends no message and changes no order.

What would have to be true to let it run unwatched. More than is true today. The independent scored exam, written by a tester and a clinical reviewer, hasn't been taken yet, so a clinician reads every flag, and an abstain is treated as unchecked, never as safe.

If you are building an AI healthcare product and want to use the MCP, email [email protected] with a line about what you are building. You'll be notified when the design sessions are complete. When you do test it, synthetic data only, never a real patient.

The meaning problem showed up all over the news this week, mostly under other names.

Pick the five nouns your product can't work without, customer, order, active, whatever yours are. Write a one-sentence definition for each. Then open your database, your interface and your main prompt, and write down the word each one actually uses.

Any cell that uses a different word, or the same word with a different meaning, is a decision somebody made without you. Fix the one your agent reads first, which is usually the prompt, and write the definition into it.

If you want to go deeper

Find the word in your product that means two things.

WHEN TO USE IT. Before you add an agent to a product that already has data in it.

WHAT IT DOES. It turns the meaning problem into a list you can argue about, one row at a time.

TIP. "Mark it as your guess" is the line that matters. Without it the model hands you definitions that sound settled, and you will paste them in as if somebody decided them.

Lex

Sunday, October 18, 5am Central. The online round of AMD Developer Hackathon: ACT III on lablab.ai closes. Teams of one to six, you join the AMD AI Developer Program to enter, and there is a Health and Wellbeing track. The finale is in Italy, travel not covered. Details and sign-up

Monday, October 26, 4pm Central. Build With AI: Basics on Devpost closes. Adults can enter as individuals, teams or organizations, no purchase is needed, prizes are $1,250, $750 and $500, and some regions are excluded. Rules and entry

Tuesday, October 27, 8am Central. Life After Code, GitLab's Transcend hackathon on Devpost, closes. Build agents on the GitLab Duo Agent Platform for everything after the code is written, and pick how much autonomy they get, approve every step, approve the outcome, or fully hands-off. Adults, solo or teams, some regions excluded, top prizes $5,000. Rules and entry

Wednesday, October 28, 3pm Central. The AWS Communication Developer Services Agentic AI Partner Hackathon closes. It is for AWS Partner Network members only, building agents on AWS messaging such as SMS and WhatsApp, which is where patient-messaging ideas would fit. Top prize $18,000. Rules

Sunday, November 8, 5pm Central. Vultr's Agent Rush on lablab.ai closes, online November 3 to 8 with an on-site day in Salt Lake City. Agents that do real work inside sandboxes, with containment required, isolation, secrets kept out, resource limits. Teams of one to six, 18 and over, $5,000 plus $3,000 in Vultr credits. Details

Thursday, November 12, 2pm Central. The PayPal AI Hackathon on Devpost closes. Build an app on PayPal's free developer sandbox with any AI tool, with a $5,000 prize for agentic commerce and $12,000 at the top. Adults, solo or teams, some regions excluded. Rules and entry

I couldn't find an in-person hackathon in Kansas City before mid-November. If you know of one, reply and I'll put it here.

Kansas City

Tool verdicts, templates and reference material live on the site at charmthirteen.com.

Decide what your product means before the tool builds it, which in clinical data is your ontology and in a front end is your brand, and that is Session One of Claude Loves Lovable, Brand Before You Build, on October 28. Register here

What's one word in your product that means two different things to two different people?

Hit reply and tell me. And if you build AI healthcare products and want the Orinyx MCP, [email protected] gets you on the list for when the design sessions are complete.

Share the newsletter: artificiallydesigned.charmthirteen.com