Roy Wiggins built a Python preprocessor called jevlang that sounds like a joke on first read: every if, elif, while, and match header in your source gets rewritten into a runtime call to a language model named Jev. You can write conditions in English — while there are bottles left: — and the condition text, the line number, and the current locals get shipped off so the model can decide whether to take the branch. The repo's own examples include FizzBuzz with English conditions, tic-tac-toe adjudicated by the model, text adventures where the model is dungeon master, and a "vibe sort" that orders foods by perceived spiciness.

What earns this an entry isn't the bit, though — it's the restraint built around it. The default backend is a fake, fully offline simulator that makes no network calls at all. There's an interactive ask mode where a human stands in for the model. There's a hard call budget so a real run can't quietly burn through an API bill. And there's a mode named, pointedly, "Luddite mode," which falls back to ordinary local Python evaluation whenever a condition is valid Python to begin with, rather than bothering Jev. Taken together, these aren't the choices of someone selling "AI writes your control flow now." They're the instrumentation of someone building an apparatus to find out, concretely, how far a program's semantics can be stretched once its smallest, most atomic decision — a single branch — becomes nondeterministic and externally negotiated.

This is exactly the place where SAE wants "already-construct" kept apart from "remainder." Phrases like "AI coding," "vibe coding," and "agentic coding" have, over the past couple of years, already been written into product launches, VC memos, and job postings — they are the settled sediment of a chisel-construct cycle that has already closed, complete with names, business models, and taking-sides. jevlang refuses to sit under that finished name. It isn't another AI coding assistant; it takes the same underlying move — substituting a model's judgment for deterministic evaluation — and pushes it to a limit where it stops being useful and becomes purely diagnostic, aimed at the smallest possible target: one if statement. It compresses an argument that mostly happens in blog posts and conference talks into something you can actually run — something that waits on a network call, costs money, and stalls when the model hedges. No existing name covers this fully: it isn't quite another entry in a list of funny esolangs, and it isn't another AI coding copilot. It's a specific, working shape given to the part of that argument that's still forming and hasn't been written down yet.

Seeing it now matters more than seeing it later precisely because it hasn't been absorbed into a stable category yet. In a few months this will most likely be filed under "funny AI-coding parodies," cited and reposted by people who never install it, never run it, never read what "Luddite mode" actually does. Right now it sits in a repository almost nobody has starred, and its README still carries a line — "modern coding style: conditionals no longer have to be Python at all" — whose tone hasn't fully decided whether it's serious. That undecided tone is the remainder still being alive. The day this gets flattened into "haha, AI can write if-statements now," the part of it that was a serious apparatus stops mattering.

Han Qin (秦汉) · @hqinjarsy · SAE Judgment & Aesthetics