Writing is cheap. Review is the work.

Writing got cheap this year. The work didn’t disappear. It moved.

A model can now produce a competent draft of almost anything — a thesis chapter, a grant section, a conference submission — inside a single conversation. The sentences are fluent. The structure is plausible. The citations are formatted to the right style. And none of that tells you whether the thing is true, whether the argument is earned, whether you would put your name under it.

In our first entry we argued that this technology is not binary — that the most honest answer to most real questions is it depends. We left one question open. Human-in-the-loop, we said, cannot mean human-in-everything; so which loop, at what moment, and with what information? Obelus is our first attempt to answer that question with something you can actually run. It is a small tool. It is open source. It does exactly one thing — and it is the clearest expression we have yet built of what we believe.

A draft is not a paper.

When you write the draft yourself, review is something that happens later, to a finished thing, performed by other people. When the model writes it, the roles invert before you’ve noticed. You are no longer the author defending a claim. You are the reviewer, deciding whether the claim has been earned — and the text in front of you is fluent, confident, and entirely unaccountable. It will not be embarrassed if it is wrong. It will not lose tenure. It does not have a name.

A draft is a response. A paper is a commitment — to a method, to a literature, to the people who will cite it and the one whose name sits in the byline. The distance between the two is precisely the work the model cannot do for you, and it is not small. Closing it is judgment. And judgment does not survive being skimmed at the speed the draft was produced. It needs somewhere to happen.

A draft is not a paper. It is a claim waiting to be judged.

Human-in-the-loop cannot mean human-in-everything.

There is a lazy way to keep a human in the loop: make the human redo the work. Read every line, rewrite every paragraph, treat the model’s output as a rough draft to be retyped. This feels responsible and is mostly theatre. It doesn’t remove the bottleneck — it is the bottleneck, now wearing a safety vest. If the price of trusting the machine is doing the whole job again by hand, the machine bought you nothing.

The useful way is narrower and harder. Find the one loop where a human’s judgment actually changes the outcome, and put the human only there. For a generated paper that loop is review. The moment is specific: after the draft exists, before your name goes on it. And the information that loop runs on is doubt — not vague unease, but doubt that has been located, named, and pinned to the exact passage that provoked it. That is what the advisor does with a student’s AI-drafted chapter; what the referee does with a conference PDF read offline on a plane; what a co-author does with a collaborator’s rendered draft. None of them is writing. All of them are deciding what to trust.

Doubt is the human’s contribution.

Medieval scribes had a mark for this. When a copyist reached a line they judged corrupt or dubious, they drew an obelus — ÷ — in the margin. Not a correction. A flag. I doubt this; look again. We took the name because the gesture is exactly right for the moment we’re in.

So that is all the tool asks of you: read the paper, and mark what you doubt. Say what kind of doubt it is — this claim is wrong, this argument is weak, cut this passage, this one needs to elaborate, the wording needs to rephrase, this part is a strength worth praising. What you do not do is rewrite it. The mechanical work — finding that passage in the LaTeX or Markdown or Typst source and making the smallest edit that resolves the doubt — is handed to a coding agent.

That division is not a convenience. It is the same architecture we build into our larger systems: separate the agent that thinks from the agent that does. In a long-running autonomous system, the thinker is an orchestrator and the doer an executor. Here the thinker is you — and the tool’s whole job is to make your thinking portable.

The handoff is a file, not a chat.

What passes from the reviewer to the agent is not a conversation. It is a file — a plain-JSON bundle that names every marked passage, its category, your note, and a precise anchor into the source: a bounding box in a PDF, a line and column in Markdown, a path into the HTML. Not a screenshot. Not “the second paragraph on page three.” A coordinate the agent can act on without guessing.

A file is something a chat never gives you: a record. You can read it, diff it, keep it in the repo beside the paper, hand it to a different agent next year. When the edits are applied, they arrive as a minimal diff you approve hunk by hunk — nothing lands without your confirmation, and anything too ambiguous to place is skipped and reported, not guessed. This is what we mean when we say governance should live inside the workflow instead of being bolted on afterward. The audit trail isn’t a feature beside the work. It is the work, written down.

An honest tool can say “I don’t know.”

We hold our autonomous systems to a standard that sounds strange until the alternative has burned you: the most valuable thing an AI can do is admit when it doesn’t know. A review tool earns that standard in small, literal ways.

Obelus runs entirely offline. The paper lives on your device, in the browser’s own private storage, and the apps make no network calls while you work — no telemetry, no analytics, nothing counted. When a paper tries to reach out for an external image or script, the tool strips the request before the page can make it, and tells you it did; you decide, per paper, whether to trust it. When a mark can’t be placed safely, the agent leaves it for you rather than guessing. None of this is dramatic, and that is the point. A tool that quietly does the wrong thing is more dangerous than one that shows you the edge of what it is sure of.

It depends.

Is the claim true? Is the argument earned? Should this sentence survive contact with a referee? It depends — on evidence you have and the model doesn’t, on a field you belong to and it doesn’t, on a name going in the byline that is yours and not its. A generated draft does not retire that judgment. It makes it the whole job.

Obelus is a small thing: a review surface in your browser and a plugin for your editor, free and open source. But it is the first place our thesis touches the ground. The machine writes; the human commits. The future is not binary — and someone still has to decide, line by line, what to stand behind.