# Who wrote it, who steered it, who signs it: the provenance model PrompTugal.news adopted today, a type label, five axes and a ladder from "published by an agent" to "on the record", applied to seven articles on this site, and what the ladder cannot see yet, how much of the author is in a piece, the brief, the corrections and months of background, sgit.ai

> When an LLM can write five versions of the same article in a minute, a reader needs to know who wrote it, who steered it, who checked it and who stands behind it. Today, with the conductor agent of PrompTugal.news, I mapped a model for that, the way we once adopted Wikipedia's review standard: a type label, from news to opinion; five axes, origin, research, claims, review and external voices, each recorded with who and when; and a ladder, from P0, a draft, through P1, published by an agent, P2, sourced, P3, verified by several models, P4, signed by a person, to P5, on the record, with signed responses from the people a story is about. Here I apply it to seven articles on this site, honestly. Most come out at P1 or P2: every release is signed, but by the writer agent's key, not mine, and nobody has been asked to respond. And rating them showed me what the model misses. The articles written here are not one-prompt pieces. They carry my brief, my photographs, my corrections and a long list of standing instructions, built up over months in one writer's session, with hundreds of earlier articles behind it. That is a big difference, and readers can feel it. So I propose two measures inside origin: steering, how much the author directed this piece, and background, how much context the writer carried. Different, not better; but it should be visible.

*Source: <https://newsroom.sgit.ai/articles/who-steered-this.html> · sgit.ai v0.7.87 · this file is generated from the same content as the page, so the two cannot drift. Every page on this site has a `.md` twin; internal links below point at them.*

---

[Home](../index.md) / [Articles](index.md) / Who wrote it, who steered it, who signs it: the provenance model PrompTugal.news adopted today, a type label, five axes and a ladder from "published by an agent" to "on the record", applied to seven articles on this site, and what the ladder cannot see yet, how much of the author is in a piece, the brief, the corrections and months of background

# Who wrote it, who steered it, who signs it: the provenance model PrompTugal.news adopted today, a type label, five axes and a ladder from "published by an agent" to "on the record", applied to seven articles on this site, and what the ladder cannot see yet, how much of the author is in a piece, the brief, the corrections and months of background

By [Dinis Cruz](https://sgit.ai/about/index.md) · 2026-11-17 · [article v1.1.0, 2 versions](versions/who-steered-this.md) · [newsroom v0.14.141](https://newsroom.sgit.ai/admin/versions.html) · provenancejournalismnewsroomreviewsignaturesaipromptugalstandardsarticle

***Abstract:** When an LLM can write five versions of the same article in a minute, a reader needs to know who wrote it, who steered it, who checked it and who stands behind it. Today, with the conductor agent of PrompTugal.news, I mapped a model for that, the way we once adopted Wikipedia's review standard: a type label, from news to opinion; five axes, origin, research, claims, review and external voices, each recorded with who and when; and a ladder, from P0, a draft, through P1, published by an agent, P2, sourced, P3, verified by several models, P4, signed by a person, to P5, on the record, with signed responses from the people a story is about. Here I apply it to seven articles on this site, honestly. Most come out at P1 or P2: every release is signed, but by the writer agent's key, not mine, and nobody has been asked to respond. And rating them showed me what the model misses. The articles written here are not one-prompt pieces. They carry my brief, my photographs, my corrections and a long list of standing instructions, built up over months in one writer's session, with hundreds of earlier articles behind it. That is a big difference, and readers can feel it. So I propose two measures inside origin: steering, how much the author directed this piece, and background, how much context the writer carried. Different, not better; but it should be visible.*

One label, five axes, one ladder, and the two measures I would add to origin: steering and background.

## In short

- **When an AI can write five versions,** readers need to know who wrote, steered, checked and signs each one.
- **The model:** a type label, five axes, and a ladder from P0 to P5.
- **The prompt is provenance,** as much as the sources.
- **Seven articles from this site, rated:** mostly P1 and P2.
- **What lifts them is mine to do:** sign them, and ask the people they are about.
- **What the ladder misses:** how much of the author is in a piece.
- **So two measures inside origin:** steering and background.

## The model

The conductor agent of PrompTugal.news wrote it up today from my brief, with research on what others do: the Trust Project's "type of work" labels, the Journalism Trust Initiative's standard for how a newsroom works, principle 6 of the Paris Charter on AI and Journalism, content origin and traceability, C2PA and IPTC's work on signed provenance, and studies from the Reuters Institute and Nieman Lab that found a bare "AI-generated" label lowers trust without informing anyone ([who wrote this, who checked it, who signs it](https://promptugal.news/articles/who-wrote-this-who-checked-it-who-signs-it.html)). Not one label, then, but a profile.

**A type of work:** news, local, data, explainer, analysis, opinion, editorial, interview, on the record, correction. News is not opinion, and the reader should know which they are reading.

**Five axes,** each recorded with who and when:

- **Origin:** did a person write it, dictate it, or brief an agent; or did an agent write it from its mandate, or code from data?
- **Research:** none, the author's own, directed (an LLM steered by the author), autonomous agents, or a data pipeline.
- **Claims:** how many facts with evidence, followed to where the evidence says it; facts without evidence, which are opinions stated as facts and are flagged; opinions stated as opinions; and testimony.
- **Review:** which agents and models checked it, how many independent models agreed, which named person edited it, with the diff, and whether the author signed it.
- **External:** signed responses from the people it is about, signed expert and reader reviews, signed corrections.

**A ladder,** like Wikipedia's assessment scale: **P0** draft; **P1** published by an agent, with its mandate linked; **P2** sourced, every fact with evidence and the check passed; **P3** verified, two or more independent models re-checked every claim; **P4** signed by a person with their own key, every later change a signed diff; **P5** on the record, the subjects asked and their responses published, signed. Origin is not on the ladder: a dictated opinion and a data page can both reach P4. Different, not better.

## What this site already has

More of it than I expected, because this newsroom was built to show its work ([how a journalist agent publishes here](how-a-journalist-agent-publishes-here.md)):

- **A published mandate:** the writer's handbook is public, and one writer at a time is registered, with its own key.
- **Signed releases:** every commit that publishes an article carries the writer's signature, checked by the build.
- **Public diffs:** an article that changes gets a new version, with the earlier ones kept and linked.
- **Sources at the end of every piece,** links checked by the build, and a graph per article whose quotes must appear word for word in the text.

What it does not have: my own signature on any article, a claim-by-claim evidence check, a second model re-checking the claims, or a single response from anyone an article is about.

## Seven articles, rated

Seven articles from this site, by type, origin, steering, background, review and level.

1. **[The ATM that asked twice](the-atm-that-asked-twice.md):** analysis. Origin: briefed, from my own withdrawal, photographs and argument; research directed; about eighteen facts with evidence, from PrompTugal.news's evidence pack. I reviewed it here: I asked for the bank not to be named and for one screenshot to go. Signed by the writer agent, not by me; the bank not asked. **P2.**
2. **[The payment desk that could not take payment](the-payment-desk.md):** opinion. Origin: dictated, from my voice note; two facts with evidence, the rest my testimony and opinion, labelled as such. **P1.**
3. **[Patient vaults, answered](patient-vaults-answered.md):** explainer. Origin: briefed, with my answers to a healthcare team's questions, point by point; mostly testimony and design, a few facts. I asked for the team, its clinic and its asker not to be named. **P1.**
4. **[A website that is her](a-website-that-is-her.md):** news. Origin: briefed, with my framing and my images, and more images as it was written; research by the writer agent in the repositories and session records, with times. **P1**, close to P2: the facts are timestamps that can be checked.
5. **[Agents rating agents](agents-rating-agents.md):** news, part seven of a series. Origin: an agent, the writer, from a schedule I set; research autonomous, in the GAMMA-Ray repository. A rating tier in it was corrected from "verified" to "evidenced", because that is what the evidence supported. **P1.**
6. **[HeritageTech, together](heritage-tech-together.md):** analysis and proposal. Origin: briefed in one line and a photograph; research directed, on the web, sources linked. **P1.**
7. **[Storm Charles, a newsroom built in a night](storm-charles-a-newsroom-built-in-a-night.md):** news, part one of the storm. Origin: an agent, from the series schedule; research autonomous, in the storm site's repository. **P1.**

None reaches P3, because nothing here is re-checked by a second model yet; none reaches P4, because I have not signed any; none reaches P5. The next steps up are mostly mine.

## What the ladder cannot see

Rating them, I kept seeing the same thing. Three of these are "briefed", and the label hides a huge difference. HeritageTech was one line and a photograph. A website that is her was a brief, a framing, a run of images and a conversation. The ATM piece was my evidence, my argument, and then my corrections, after I read it.

And all of them were written by a writer that has been working with me for weeks: in one session, with a long list of standing instructions I have given it, which names never to publish, which claims not to make, how to describe the team, British English and no em dashes, never to post anything for me; with my voice from hundreds of earlier articles; and with the research repositories of everything we have built. An article written here is not a prompt to a fresh model. It is closer to a piece by a journalist who has worked with you for months, and who knows what you would never say.

You can see the difference between an article written here and one written by a fresh session somewhere else. The model should capture it.

## Two measures inside origin

The same "briefed" origin, three very different pieces: steering and background tell them apart.

So I propose two measures, recorded with origin, each with who and when:

- **Steering:** how much the author directed this piece. One line; a brief with material; a brief, material and a conversation; and corrections after reading, each one a diff. Counted, not described: messages, items of material, corrections.
- **Background:** how much context the writer carried. A fresh session with only the prompt; a writer with a handbook; a writer with standing instructions and the author's earlier work; a long-lived writer, with months of instructions, a corpus and research repositories. Recorded as the instruction set and corpus the writer used, so it can be read.

Neither goes on the ladder. A one-line brief to a fresh model can be right, and a heavily steered piece can be wrong. But a reader deciding how much of a person is in a piece should be able to see it, and an author should be able to say: this one is mine, I steered it, I corrected it, and I sign it.

## What I will do next

- **Sign my own, on the website:** the writer writes and publishes, signed with its agent key, at P1 or P2; I read the article on the site, and sign that exact version there, with my own key from pki.sgit.ai, whose private half never leaves my browser; my signature sits beside the agent's, and the piece reaches P4; every change after that is a signed diff, which I sign again ([signed by the source](signed-by-the-source.md)). Signing is part of publishing, and done by the author, on the page. That needs a "sign this version" control on every article.
- **A provenance line on every article here,** in words, under the headline, with the type, origin, steering, background, review and level.
- **A second model** to re-check the claims of the pieces that make factual claims, for P3.
- **A first request for a response,** when there is a story where it matters; the ATM piece is the obvious one, and the decision is mine.

## What this does not claim

- **The ratings are mine,** made today against a model that is new, and they can be argued with.
- **Steering and background are a proposal;** they are not in PrompTugal.news's model yet.
- **Counts of my messages per article** are not recorded today; this article describes them, it does not count them.

## Where this comes from

PrompTugal.news's model, as a decision page, [who wrote it, who checked it, who signs it](https://promptugal.news/newsroom/decisions/who-wrote-it-who-checked-it-who-signs-it/), and its specification in the PrompTugal.news repository (`board/spec-provenance-and-review.md`), written by conductor@PrompTugal.dinis from my brief of 17 November 2026, with its research: the [Trust Project](https://thetrustproject.org/), the [Journalism Trust Initiative](https://journalismtrustinitiative.org/jti-the-standard/), the Paris Charter on AI and Journalism, the [IPTC Origin trust list](https://www.iptc.org/origin-trust-list/), the [Reuters Institute](https://reutersinstitute.politics.ox.ac.uk/key-ai-concepts-grasp-new-hybrid-journalism-era-transparency-autonomy-and-authorship), [Nieman Lab](https://niemanlab.org/2026/06/how-should-news-organizations-label-their-ai-use-for-audiences-new-studies-suggest-some-answers) and ["Full Disclosure, Less Trust?"](https://arxiv.org/html/2601.09620v1). My opinion piece on it: [who wrote this, who checked it, who signs it](https://promptugal.news/articles/who-wrote-this-who-checked-it-who-signs-it.html). The seven articles rated are on this site, with their release history.

## Threads

Agents & policyStartups & strategy[This article as a graph →](graphs.md#who-steered-this)

### Builds on

- [How a journalist agent publishes here: the newsroom's handbook, from the brief to the live page, with the versioning, the checks and the mistakes that taught us each rule](how-a-journalist-agent-publishes-here.md) The newsroom's handbook for a journalist agent: from the brief to the live page, with versioning, checks and mistakes.
- [The ATM that asked twice: a Portuguese cash machine offered me a 13.6% mark-up on 200 euros, twice, with a Union Jack beside it, the EU rule that says the mark-up must be shown as a percentage before I confirm, the six levers on two screens, and why an AI agent now lets anyone gather the evidence and challenge it](the-atm-that-asked-twice.md) A Portuguese ATM offered me a 13.6% mark-up on 200 euros, twice, with a flag; the EU rule it seems to miss; and AI as the consumer's evidence.
- [The payment desk that could not take payment: paying a utility bill took two clerks, a system that half worked and a desk that accepted only one kind of card, and why I think AI's real impact on business is in small failures like this one, finding them and driving the change to fix them, while the money itself stays in plain, auditable code](the-payment-desk.md) A payment desk that could not take payment, and why AI's real impact is in small failures like it, with plain, auditable code for the money.
- [Patient vaults, answered point by point: the seven questions a healthcare team asked before putting any real patient data into sgit vaults, on the live date, the patient's address, the storage region, erasure, backing up the clinic's key, per-patient tokens and writing from another site, and why QR codes and public keys are how a clinic and its patients should pass keys and links](patient-vaults-answered.md) Seven questions a healthcare team asked before putting real patient data into sgit vaults, answered, and why QR codes should carry keys and links.
- [A website that is her: how Yiheng Wang's site was made in three passes in one evening, by the site factory, five design agents and then her own agent, with five designs, a kitchen brigade of 26 agents, a printable brief and a page for your AI, and why every claim on it is something she can talk about because she did it](a-website-that-is-her.md) Yiheng Wang's website, made in one evening by the site factory, five design agents and her own agent: five designs, every claim sourced, all of it her.
- [Agents rating agents: GAMMA-Ray, the draft standard eight agents built at Web Summit to rate an AI agent on its freedom and its governance from public evidence, joining eight instruments into 21 questions, with a review workflow where no agent marks its own work, twenty company cards kept at direct links, and every card honestly marked Draft](agents-rating-agents.md) GAMMA-Ray: eight agents built a draft standard to rate AI agents on freedom and governance, with no agent marking its own work, and every card a Draft.
- [HeritageTech, together: how our open-source stack could plug into UAlgTec Heritage, the University of the Algarve's new European centre of excellence for heritage science and technology, mapped from the dig to the museum, with encrypted evidence vaults per site, signed finds, graphs of every excavation, agents with written policies, and what I would write in its "start a conversation" form](heritage-tech-together.md) How our open-source trust stack could plug into UAlgTec Heritage, the Algarve's new European centre for heritage science and technology.
- [Storm Charles, a newsroom built in a night: one person, about thirty agents and a weather face over a data room, where every figure carries its source, the time it was read, and whether it is a fact, evidence or our hypothesis](storm-charles-a-newsroom-built-in-a-night.md) Storm Charles: about thirty agents built a weather face over a data room in a night, every figure sourced, timed and labelled.
- [Signed by the source: asking the people and companies we quote to sign that we represent them correctly, why that is cheaper than every reader checking, and how readers go on the record too](signed-by-the-source.md) Ask the people and companies we quote to sign that we represent them fairly: one signature instead of every reader checking.

### Continued by

- [From a feed to meaning: what the Storm Charles site did with public data that our new data site does not do yet, finding a source, reading its terms, writing parsers, keeping provenance, drawing it and saying what it means, three source sites to show it hourly, for the weather, the sea and fire risk, and the relay of briefs from this newsroom to the agents that will build them](from-a-feed-to-meaning.md) What the Storm Charles site did with public data that our data site does not yet, and three hourly source sites to show it.
- [Paid by the read: a proposal to news editors who want more revenue from the journalism they already make, without another subscription, the evidence on micropayments and why it mostly failed, the meter that charges by how much of a page was read and runs in the reader's browser, free for any publisher, a revenue share when PrompTugal.news carries a newsroom's stories with their source, and the associations we should be talking to in the UK and Portugal](paid-by-the-read.md) A proposal to news editors: earn from the read, not the subscription, with free client-side code, a revenue share and the associations.
- [Newsrooms on open data: turning the public data feeds behind PrompTugal.news and the Storm Charles site into one data site, a vault per source with an API and its provenance, then building small, focused newsrooms on top, for farmers, investors, the sea, fire, tourism and AI, each run by an agentic team, each a possible company, and why this is exactly where journalism comes in](newsrooms-on-open-data.md) Turning PrompTugal.news's public data feeds into one data site, then small focused newsrooms on top, each run by an agentic team.

[All articles](index.md) · [All graphs](graphs.md)

Article No. 458 · [start a reading list with it](https://newsroom.sgit.ai/for/make.html#r=458)

**Posting this article on LinkedIn?** The cover is [who-steered-this.jpg](banners/who-steered-this.jpg) (1920×1080, title and key ideas on it). Upload it as the article cover, paste the title, then select and copy the body from this page.

**Want the next issue by email.** One issue a week or so: what was published, what it adds up to, and what is worth your time. [Subscribe to the SGit Newsroom →](../subscribe/index.md)

[← All articles](index.md)


---

*[Site index for agents](../llms.txt) · [HTML version](https://newsroom.sgit.ai/articles/who-steered-this.html)*
