Backticks

In a Jev question, a backtick is a pointer. TypeSafe’s documentation uses it to aim a question at one part of the state: a backticked path like `support.tickets[0].message` means judge this field. Anyone who writes markdown uses backticks for something else, anything that looks like code: a command, a flag, a filename.

The two habits collide in a frame written the markdown way. A triage design for Omarchy, drafted by an AI agent, wrapped Omarchy’s own command names in backticks seventeen times, and none of them named anything in the state. We measured what that does. On five reports that plainly did or did not describe a command failing, the stray backtick changed nothing past the second decimal. On the one report that described the failure without naming the command, it dropped the answer from 0.87 to 0.77, five times the widest spread between repeats. The rule we now enforce is mechanical: a backtick names a state field, and everything else goes in plain quotes.

Story explains it with the defined terms in a lease. Technical has the six reports, the three wordings, and what the measurement cannot say yet.

Read a lease and you will find words with capital letters in the middle of sentences: the Tenant, the Premises, the Term. Each one points back to a definition near the top of the document. When a clause says the Tenant must keep the Premises clean, a careful reader does not guess what “Premises” means. They look it up, and the clause means exactly what the definition says, no more and no less.

Now imagine a lease where the writer also capitalised words for emphasis. “The Tenant must not make Loud Noise after ten.” A careful reader goes looking for the definition of Loud Noise and finds none. What do they do? They read the clause more narrowly than the writer meant, as if Loud Noise were a specific, defined thing, and the neighbour’s ordinary shouting might not count.

Jev is that careful reader. Every question it answers is asked about a set of facts your code hands it (the state), and TypeSafe, the company behind Jev, uses a pair of backticks around a name to mean: this exact field of the state. The backticks are the capital letters. The fields of the state are the definitions.

People who write software put backticks around all sorts of things, because in the documentation and chat tools we use every day, backticks mean “this is code”. A command name, a file, a flag, all backticked out of habit. AI models learned to write from the same documents, so they do it too.

We found the collision in a triage design for Omarchy, an open-source Linux desktop, drafted by an AI agent. Its questions wrapped Omarchy’s own command names in backticks seventeen times. None of those names was a field of the state. In lease terms, seventeen capitalised words with no definition anywhere in the document.

The question was whether it mattered, so we measured it. We wrote six short issue reports and asked Jev one question about each: does this report say the update command fails? Once in plain words, once with the command’s name in backticks, and once with backticks around the real field that held the report.

On five of the reports, the backticks changed nothing worth mentioning. Where a report plainly named the command and quoted its error, every wording agreed it failed. Where the failure belonged to a different command, every wording agreed it did not.

The sixth report said “the update command errors out halfway through every time I run it.” It never used the command’s actual name. Asked in plain words, Jev put the chance it described the failure at about 0.87. With the name in backticks, 0.77. It looks as if the careful reader went looking for the exact defined term, did not find it written in the report, and read the question more narrowly, the way the lease reader treated Loud Noise.

That is the part worth holding onto. The stray capital letters did no harm where the document was clear. They changed the answer exactly where the document was ambiguous, which is the only place an answer was ever in doubt.

So the rule in our frames is the lease-writer’s rule. Capitalise only defined terms: backticks go around the names of fields in the state, and nowhere else. Anything else that deserves quoting, a command, a flag, an error message, goes in plain quotation marks. And because habits come back the moment nobody is watching, a checker now reads every question before it can run and refuses any backtick that points at nothing in the state (a dangling reference, in the trade’s words).

One character carries two conventions. TypeSafe’s documentation uses backticks inside a question’s instructions to reference the state: “Use a backticked dot-and-index path to point a question at a specific nested value, such as `support.tickets[0].message`.”¹ Markdown, and every chat and documentation tool built on it, uses the same character to mark anything code-shaped. An independent field guide on pairing Jev with LLMs that circulated the same day follows the pointer convention, “Does `llm_output` directly address `request`?”, though it is not a TypeSafe document and says so.² Reading the docs through on a recording for this work, I called the backtick the place “where a lot of the conceptual work needs to be done if you’re used to working in API land.”³

Where they collided

Two AI agents designed an issue-triage system for Omarchy from the same brief. Checked after the fact, one design’s questions contained seventeen backticked tokens that named nothing in their state: Omarchy command names and a package-manager flag quoted from issue reports. Every backtick in the other design named a key its state builder actually sends.⁴ The difference looks like habit rather than design: the method one of the agents was following did not teach the rule at the time.

What a stray backtick does

The measurement holds the state fixed and varies only the wording.⁵ The state is a project description plus issue.body, one of six synthetic reports. The question is a Noul, asked three ways:

• plain: “The report says that the omarchy-update command fails.”
• stray: “The report says that the `omarchy-update` command fails.” The backticks point at no field.
• pointer: “The report in `issue` says that the omarchy-update command fails.” The backticks point at the field that holds the report.

Each wording was asked three times per report, 54 calls in all. Ranges are across the three repeats.

Report	plain	stray	pointer
names the command, quotes its error	0.97	0.97–0.98	0.97–0.98
the command succeeds, a different one crashes	0.02	0.02	0.02
unrelated problem, command not mentioned	0.02–0.03	0.03	0.02
failure shown in a code block	0.97	0.97–0.98	0.97
describes the failure, never names the command	0.86–0.88	0.76–0.78	0.83
the command ran, a later problem followed	0.04–0.05	0.05	0.03–0.04

On five reports the stray backtick moved nothing beyond the second decimal. On the report that paraphrases the command, it cut the probability by about 0.10, against a spread between repeats of 0.02 at most. The likely reading, which this data cannot prove, is that a backticked name is treated as an exact reference: where the report contains the literal string, nothing changes, and where it describes the same thing in other words, the answer leans toward no. That reading is inferred, not measured.

The pointer wording lowered that answer too, to 0.83. Narrowing a question to one field is a change of meaning as well, and it should be chosen for that reason rather than added as decoration.

The rule, enforced

In our frames a backtick names a state path and nothing else: a top-level key, a nested path, or a list item. Commands, flags and error strings go in plain quotes. A checker reads every question’s instructions and criteria before a frame can run, resolves each backticked token against the frame’s declared fields (nested paths, and lists written with or without an index), and fails the frame on any token that resolves to nothing. The same check, run over the two agent-drafted designs, produced the seventeen.

What would falsify this

The measured effect rests on one ambiguous report, six synthetic states and one model version. It would be falsified by a larger set of reports that paraphrase the thing a question names, where stray backticks move answers no more than repeats do; the rule would then be hygiene rather than accuracy. Two related claims are untested: that a nested pointer such as `issue.comments[0].body` actually confines attention to that field, and that the effect holds on later model versions.