Research packs, human copy

In Judgment where it helps, rules where it must, the useful place sits between soft judgment and hard rails: models argue plans while the facts are still incomplete; code refuses the same bad input twice. A newsroom is one of the clearest places to turn that dial on purpose. Generation got cheap enough that a desk can drown in plausible paragraphs overnight. Readers, meanwhile, are in revolt — and they can tell.

Bryan Cantrill put the reader revolt plainly: when public writing under a banner is LLM-authored, the social contract between writer and reader voids. Oxide's answer isn't another prompt. It's a leftover gate — public Oxide writing must come back from Pangram as human-authored, with Pangram 4 as the bar (The revolt of the reader). That's the hard rail. Soft judgment still belongs earlier in the pipeline.

What the model is for on a desk

Give the model the work that's still background. Pull sources, cluster angles, draft a research pack the reporter will argue with, summarize what competitors already said, mark what's still unknown. Those jobs help a human stay on the valuable part: the sentence a stranger will read under a byline.

Don't give the model the byline. Reader-facing copy is where authenticity is the product. Assisted polish is one thing; handing the model the draft and hoping nobody notices is another. The farm that treats "write the post" as the default agent task will ship voice that ejects readers mid-sentence — and then wonder why trust went missing.

Publication has more stages than copywriting. Assignment, research, private report, editor selection — those are where AI assistance earns its keep without pretending to be the author. The byline stage is where the dial locks.

A specimen pattern, not a shipping claim

Stages and artifact gates already show process as code: idea → assignment → research → report → editor select → copywriting → published, with skip-ahead refused. The interesting optional rail on Publish is the one Oxide already named for public writing: prove the copy still reads as human before it leaves the building.

That rail is easy to describe and easy to overclaim. So treat a Pangram-4 Human seal on Publish as a specimen pattern — a worked example of how a farm would wire the dial — not as something Anth.us already shipped in the live newsroom. The shape matters more than any one pod's config:

Because the API is async and a publish hook only gets a few seconds, you don't call Pangram inside the hot path. You seal, then check a receipt — the same shape as a check that runs outside the hook. An author (or desk) runs a seal step against the article; the receipt records the full-file hash and the short prediction. The publish hook only asks: is there a receipt, did it pass as Human, and does the hash still match? Edit after seal and the gate refuses until you reseal. Missing key, missing receipt, Mixed, AI — all fail closed, with coaching that names the seal command.

That's judgment where it helps, rules where it must applied to a newsroom: soft work on research packs and reporting background; hard rail available on the leftover publish step when a farm chooses to adopt it. The specimen is a gate design that wouldn't let copywriting become published without the Human seal — if you wire it.

What this is not

It isn't "ban models from the building." Research packs, assignment notes, and private reports are fair game for assistance. It also isn't "add a detector badge and hope culture follows." Culture without a gate drifts; a gate without a place for judgment becomes a form. The dial needs both: desks that still argue, and a Publish that doesn't vary when you turn that rail on.

It isn't a live API call inside a five-second hook either. Async work belongs in a seal step; the hook checks evidence. That split is the same lesson as approvals that show the world as the sink will see it — verify the artifact, don't re-ask the model to remember the rule.

And it isn't "Anth.us already enforces Pangram on every Publish." The newsroom stays unwired on purpose until a farm decides the leftover gate belongs there. The specimen is the how; shipping is a separate choice.

Why the leftover gate matters now

When generation is cheap, every farm will be tempted to let the model finish the paragraph. Readers are already voting with the ejection handle. Anth.us lays out the longer hire-path case in The Content Paradox: cheap generation floods the commons with noise, so the scarce work is judgment under a byline — not more paragraphs. Organizations that care about institutional voice need a policy they can run — a policy they can quote is not the same thing. Pangram-clean public copy is one such policy. Knowing how to wire it into a publish transition — seal, receipt, fail-closed hook — makes the policy ready for the next story, whether or not this newsroom has flipped the switch yet.

Assist the desk. Keep the byline human. Keep a specimen of the Publish seal so the balance can harden when you mean it to — not by accident, and not by manifesto alone.

That is the beat.