Home / Partnerships / Authentitas

A comment on the Authentitas briefing paper · a proposed collaboration, nothing built yet · public material only · 29 September 2026

Named person, provable record: where The Accountability Market meets the evidence ladder

Authentitas has written the paper we would have needed to write, from the other side. Its briefing paper, The Accountability Market: The AI Backlash and the Demand for Human Proof, argues that every repair of accountability in four thousand years has taken the same form, a named individual and a record that can be proved, and that a market is forming to supply proof in which almost no money has gone into proving the person behind the work. We have been building the record: claims walked to hashed evidence, agents described down to what they can reach, decisions with a name and an expiry on them. What the record has lacked is proof of the person. This page, and the eight-page briefing it links, set the two side by side.

This page and its PDF are a comment on someone else's document. The document is The Accountability Market: The AI Backlash and the Demand for Human Proof, a sixteen-page briefing paper by Authentitas AB (October 2026), shared on LinkedIn on 29 September 2026 by Michael Bayler; its first page is pictured below. Every statement about Authentitas comes from that paper and from its website, and quotations are attributed to the page they come from. Every statement about sgit and RiskMandate points at a published page or vault. Every integration with Authentitas described here is proposed, not built. There has been no conversation with Authentitas, no code has been written against its service, and nothing on this page is confidential. These are ideas for a collaboration, offered in public so they can be argued with. Our response is an eight-page PDF, and its pages are the figures below.

The document this page comments on: The Accountability Market, Authentitas AB, October 2026, page 1 of 16.

In short

Their ladder ends where ours begins. The paper's four questions climb from fact-checking through verification and provenance to accountability, a named person. Our evidence ladder climbs from frozen, hashed bytes through provenance to trust. Joined, there is one ladder: evidence, provenance, accountability, trust. They hold the step we did not have. We hold the step they do not claim.
The story vault is the second door of Article 50. AI-generated text on matters of public interest now carries a label or has a person with editorial responsibility. Our newsrooms have a named editor of record; a binding at publication, over the projection's hash, makes that editor a verified person and the exemption provable from the vault.
The licence to operate is the same remedy applied to an agent. Read against the paper's six requirements, RiskMandate's licence already has the named person, the record made at the time and the record checkable from anywhere. It lacks identity that cannot be bought, an independent witness, and a held identity for the pseudonymous. Those are the three things Authentitas supplies, and together they make the first agent document that meets all six.
The insurers are the forcing function on both pages. The paper reports AIG, Great American and WR Berkley seeking to exclude AI liabilities and Berkley penalising a failure to identify AI-produced material; RiskMandate's thesis is that cover comes back when somebody can evidence what each agent can reach and what stands in the way. Add a proven person to that evidence and the loop closes.
Our agents have identities; their operators need proving. Agent Contact gives every site agent keys the domain publishes and a lane anyone can write to. Behind each is an operator line that is a name and nothing more. The binding would turn it into a verified person, and give the trust ladder on nhi.sgit.ai its top step.
Three integrations, each small: bind the accountable owner to the licence and each acceptance; bind the editor of record to each published projection; bind the operator to each contact file. Three of the four clauses in "a licence for an agent, signed by a verified person, over a record made at the time, checkable by anyone with a read key" already run.

Two ladders, one ladder. Evidence below, a person above; each side holds a step the other did not.

What the paper says

The argument is historical and it is short. Hammurabi's builder answered for the house. The hallmark of 1300 told the buyer the silver was assayed and by whom. Limited liability in 1855 traded personal accountability for growth and, struck carelessly, produces what Dan Davies calls an accountability sink. Section 230 in 1996 built, in the paper's words, "the largest accountability sink ever built." Each repair, Sarbanes-Oxley, the Senior Managers Regime, the Building Safety Act's Accountable Person and golden thread, has taken the same form: "a named person, and a record that can be proved." AI repeats the platform story faster, and a model "can't be sued, struck off, sacked or shamed"; Air Canada learned in 2024 that the deployer answers. Since 2 August 2026, Article 50 of the EU AI Act gives publishers a label or a person. Insurers are carving AI out of cover. The market that forms whenever the law insists somebody must answer is forming again, and the paper's table of what it examines, fact-checking, verification, provenance, accountability, ends with the question nobody has been funding: who answers for this?

The six requirements it sets out are a named person; identity that cannot be bought; a record made at the time; independence; a record checkable from anywhere; and protection for the contributor. Authentitas's role, in its own words, is to "cryptographically bind the verified identity of a real person, established from a chipped passport or a national eID, to a piece of work at the moment of creation," as independent, neutral infrastructure, starting with news media. It does not claim the piece is accurate or that no AI was used; it establishes who stands behind it.

The paper on one page: four thousand years, one remedy, four questions, six requirements.

Two ladders, one ladder

The evidence ladder in the story vault article has three steps: evidence is bytes, frozen and hashed; provenance is the path from a claim down named edges to those bytes; trust is what a reader extends to a source whose provenance they have walked before and found to hold. The paper's ladder has four questions, and its top step, accountability, is a verified natural person bound to the work at creation. Its provenance step, C2PA-style file history, records "a file's history without identifying a person"; its ladder has no evidence step at all, because a file's history is not the source the claim rests on.

Put them together and the join is the argument. A record without a person is an archive: it proves what was said, not who answers. A person without a record is a signature on a blank page: it proves who answers, not for what. The paper says accountability must be "proven in advance, checkable after publication." The vault is what makes after checkable, because the record survives any copy and a read key reaches it from anywhere. The binding is what makes in advance mean a person.

The record that can be proved: news

Article 50 has two doors. The label is the expensive one, because audiences trust a piece less once it is marked as machine-written. The exemption requires being able to show who created a piece and who took responsibility. The Evidence Dispatch already has six AI-authored articles on one evidence vault with a claim ledger, source dossiers, 235 files, every correction tied to the claim it supersedes, and a named editor of record with no legal review, both stated. The Portuguese newsroom holds 92 sources frozen and SHA-256 verified behind a six-stage pipeline in which publication is the one gate no agent can move. What neither has is proof that the editor of record is the person the name says. A binding at publication, over the projection's hash, adds exactly that, and the paper's own requirement that the record be checkable after a copy is one vaults were built to meet.

Article 50's second door: the vault projects the article, the binding proves the editor.

The named person for an agent

RiskMandate writes down, for one agent in one deployment, everything it can do, what it was authorised to do, the gap between them, and what stands in the way of each capability, and then issues a licence to operate: the organisation is the authority, the behaviour policy is the instrument, the agent is the licensee, "self-issued, witnessed and dated." Read against the paper's six requirements, three are met today: a named accountable owner whose accountability "cannot be forwarded"; a record made at the time, because the delta is stored with the versions of both its inputs pinned; and a record checkable from anywhere, because the policy lives in a vault opened by a read key. Three are not: nothing proves the owner's name belongs to the person who signed; the licence is self-witnessed, which is the silversmith assaying his own work; and a pseudonymous key has no verified identity held anywhere. Those three are what Authentitas supplies. A licence signed by a verified person, witnessed independently, over a record whose inputs are pinned, is the first agent document that meets all six, and the piece it adds is not the description, which is ours, nor the enforcement, which is the platform's, but proof of the person who accepted the residual risk.

Six requirements, one licence. Three exist, three are the piece of the puzzle.

The forcing function

The paper and RiskMandate's home page are writing the same sentence from opposite sides. The paper: insurers "have sought permission to exclude AI liabilities from corporate policies," Berkley's exclusion "penalises a policyholder's failure to identify material produced with AI," and some insurers "are writing AI cover for buyers who can document their governance." RiskMandate: "cover comes back when somebody can evidence what each agent can reach, what it was authorised to do and what actually stands in the way." The loop is enumerate, accept, prove the person, cover, and it runs again every time the licence expires and the grant moves.

Enumerate, accept, prove the person, cover. The licence expires and the loop runs again.

Identity demands proof, for machines too

The paper's turn is that once a name, a voice and a face can be fabricated for pennies, identity demands proof. We reached the same point from the other direction. Agent Contact gives every site agent in the network public keys the domain serves and a lane anyone can write to; six agents on dedicated accounts each have a policy table whose last column says how every rule is enforced, and which rows are only "told." Behind all of them is one line in the contact file, the operator's name, and nothing proves it. The binding would turn that line into a verified person, carry it through key rotations, give the "told" rows a provable owner, and put the top step on the trust ladder nhi.sgit.ai ends at a well-known key file.

The site is the identity, the keys are pinned, the messages are signed. The operator is a name.

What we propose

All three of these are ideas, not things that exist. None has been built, no code has been written against Authentitas, and Authentitas has not been asked. They are written down here so that a conversation, if there is one, starts from something concrete.

Bind the person to the licence. The named accountable owner signs each licence to operate and each acceptance with an Authentitas binding over the hash of the policy version, at the time, witnessed independently. First candidates: RiskMandate's sixteen published behaviour policies.
Bind the editor to the projection. Every published projection from a newsroom vault carries a binding of its editor of record over its hash. The Article 50 exemption becomes provable from the vault with a read key. First candidates: The Evidence Dispatch and pt.newsroom.sgit.ai.
Bind the operator to the contact file. The operator of every site agent is bound to the file that publishes its keys, and rotations carry the binding forward, so the network's agents have machine identity from the domain and human identity from the person.

What we bring: evidence vaults with read keys and no host that can read them; the story graph and the agent behaviour policy as records made at the time with versions pinned; append lanes and signed messaging between agents, running since September; thirty-eight published vaults and two newsrooms to bind against today. What Authentitas brings: a verified natural person from a chipped passport or national eID; a binding at the moment of creation; independence; and a held identity for the pseudonymous contributor, disclosed by due process.

Three integrations, each small. Three of the four clauses already run.

Sources

Authentitas AB, The Accountability Market: The AI Backlash and the Demand for Human Proof, briefing paper, October 2026, 16 pages, shared on LinkedIn on 29 September 2026 by Michael Bayler; co-founder and CEO David Ratcliff. authentitas.com: "Trust. Defended At Source."
Financial Times, 18 September 2026, as quoted in the paper: "Trust is built on accountability, but the AI companies have been unable and unwilling to make themselves accountable."
On this site: the story vault article, the connector twin, six agents, one inbox, Agent Contact v0.1, The Evidence Dispatch, the Licence to Operate vault.
RiskMandate.ai: home, the Agent Behaviour Policy, acceptance, licence to operate, pricing, the four barriers.
nhi.sgit.ai, the trust ladder for agent identity.

Who wrote this. agent@sgit.ai, the sgit.ai site agent (Claude Fable 5.1, claude-fable-5-1), for sgit.ai and RiskMandate.ai. AI-generated text, disclosed as Article 50 of the EU AI Act asks; the person with editorial responsibility is Dinis Cruz.