The accountability architecture for AI-written publications

1. Disclosure

The first component is the easiest to implement and the most consistently refused. Disclosure means the publication declares, on every piece, that an AI system drafted the prose, that a named human signed it off, and that the human is accountable for the claim it asserts. No second clause. No hedging. The byline carries the declaration; the byline is not optional.

The cost of non-disclosure is by now a matter of public record. Sports Illustrated was found in late 2023 to have published AI-written product reviews under fabricated human bylines, complete with fabricated headshots, and the story metastasised because the discovery was made by readers, not by the publication. CNET ran a parallel disaster the same year: dozens of AI-drafted finance articles published under a generic staff byline, half of which were quietly corrected after factual errors were surfaced externally. Neither publication recovered the trust it lost. Both could have inoculated themselves against the entire category of damage by saying, in advance and on every piece: this was drafted by AI; this human is accountable for it.

Disclosure is the precondition under which the remaining five components can do their work. A reader cannot evaluate the accountability of a publication that pretends its content is human-written. Verdict tracking on a hidden author is theatrical. A retraction log signed by a fictional staffer is decorative. Primary-source pinning on a piece whose authorship is itself misrepresented is the inverse of trust-building. Until the publication says how the piece was made, nothing else it says about the piece is structurally credible.

At agentmodeai.com the disclosure is on every article header, in every byline, in the footer, and on a dedicated page describing the production model in detail. The byline reads “Written by Claude · Curated and signed by Peter”. The signatory is a named human; the prose is AI-drafted; the arrangement is the publication’s public-facing fingerprint. I am the signatory of this site, and this piece argues for the architecture that the publication itself implements, that self-reference is acknowledged here once and is not what the argument rests on.

The refusal of disclosure across the AI-writing-AI publishing landscape is a commercial choice, not a technical one. Publications fear that readers will discount AI-drafted prose on sight, so they hide it. The architecture described here works the other way: explicit disclosure plus the remaining five components produces more reader trust than non-disclosure plus none of them. The empirical bet of this publication is that the second combination beats the first on a 24-month horizon.

Disclosure is necessary. It is not sufficient. The next five components make it operational.

2. Claim isolation

Every assertion of fact in an article is a claim. Claims have lifecycles. Some are true on the day of publication and remain true; some are true on publication and become wrong as the world moves; some turn out to have been wrong all along. Traditional publication does not address claims individually. The article is the smallest unit of trust the format ships. When one paragraph turns out to be wrong, the publication can correct the whole article, retract the whole article, or quietly edit the whole article, none of which is granular enough to tell a careful reader what specifically changed.

The first move of an accountability architecture is to make the claim itself the addressable unit. Every claim of fact gets a stable identifier at drafting time. The identifier is registered in a public registry. The identifier is cited inline in the article body. Other articles can refer to the same identifier when they need to lean on the same claim. External work can cite the identifier with the confidence that the reference will be stable across rewrites of the surrounding prose.

This publication uses the prefix scheme AM-NNN for parent-register pieces, OPS-NNN for the operators register, RES-NNN for downloadable tools, and AM-CANON-NNN for canon pieces of which this article is the first. The numbering is the publication’s own; the discipline is what matters. Each identifier resolves to a YAML file in a public directory of the site. Each YAML carries the claim text verbatim, the verdict at the time of last review, the date the claim was reviewed, the date of next review, and an append-only correction log. The claim text never changes after publication. The verdict mutates as evidence moves; the original sentence does not.

The closest precedents in human publishing get part of the way to claim isolation and stop. FactCheck.org publishes per-claim cards, which solves the addressing problem inside one organisation but does not extend the addressing to the underlying article it is checking. Retraction Watch maintains a public database of retracted academic papers, which solves the visibility problem at the article level but not at the per-claim level. Nature operates a retraction-notice system with stable URLs and dated notices, which solves both at the article level but not at the level of individual assertions inside an article that survives. The combination of all three approaches, applied to the per-claim grain and consumer-grade in availability, is the architecture’s contribution.

There is a structural reason this is easier inside an AI-drafted publication than inside a human-authored one. Drafting via a large language model produces explicit, separable claim statements more reliably than rhetorical human prose. Human commentary tends to blur claims into the structure of the argument, a clause here, a hedged assertion there, a footnote that quietly carries a load-bearing statistic. AI-drafted prose with the right system prompt produces sentences that say what they assert, where they assert it, and what they are willing to be wrong about. Authoring discipline that human writers reach for after a decade of practice is, for an AI-drafted publication, available on day one if the editorial instruction calls for it.

Claim isolation is the addressing layer of the architecture. Every component below it depends on the fact that the publication has a stable way to refer to what it has said. Without that, verdict tracking has nothing to track, retraction has nothing to point at, primary-source pinning has nothing to anchor, and review cadence has nothing to revisit. The address is the load-bearing primitive. Everything else is a system built on top of it.

3. Verdict tracking

A claim’s truth status is not binary, and it is not static. The architecture’s verdict vocabulary is four-valued: Holding, Partial, Not holding, Retracted. Each is a named state. Each state has explicit conditions under which it transitions to the next.

Holding means the claim still survives current evidence after the most recent review. The publication has re-checked the underlying sources, the world has not moved against the claim, and the publication is willing to continue asserting it. The next review is scheduled.

Partial means the claim was overstated. The core assertion is still defensible, but one substantive part of it has been qualified, narrowed, or revised. The correction is dated and named. The original claim text remains visible; the verdict shifts; a reader can see both the original statement and what changed.

Not holding means the claim is no longer supported by current evidence. The publication has been overtaken. The article stays online; the claim stays in the ledger; the verdict signals that the publication no longer stands behind the original statement. The original prose is preserved; the verdict is the public update.

Retracted means the claim was wrong on the day of publication. The original prose is preserved exactly as published. A dated retraction notice is appended. The corrected reality, where it can be stated, is stated. The retraction is logged in a public retractions surface that lists every retraction the publication has ever issued.

Verdict tracking is the temporal layer of accountability. A traditional publication writes a claim at time T and the claim stays at time T forever, even when the world at time T plus one year would no longer support it. A publication running the architecture commits, in advance and in public, to revisiting its own claims on a published cadence and announcing the result. The commitment is the trust contract.

A verdict changes through one of three triggers. The first is scheduled review on the cadence the publication declared at the time of publication: typically every ninety days for the parent register, every thirty to forty-five days for the operators register, where tooling pricing and SMB-relevant facts move faster. The second is reader-submitted evidence, surfaced on the corrections page, which gives an external party a structured way to ask the publication to re-examine a specific claim. The third is proactive editorial discovery: the publication finds its own error before any reader does. Every transition is dated. The history of each claim is public, visible at the claim’s permalink, and is part of the publication’s track record.

The closest precedents in human publishing each carry one piece of this and miss the rest. Peer-reviewed journals operate post-publication commentary mechanisms, letters to the editor, corrigenda, formal comments, which capture the idea that a claim’s status evolves, but not on a scheduled cadence and not with public verdict labels a non-academic reader can parse. Wikipedia maintains article-talk histories that capture the deliberation over what an entry should say, which is closer in spirit, but the talk page is not a primary-claim ledger and is not consulted by most readers. The IETF maintains errata for published RFCs, which is the closest analogue at the per-claim grain, with the dated update and the public addressability, but RFCs are technical specifications, not editorial commentary, and the cadence is reactive rather than scheduled. The architecture borrows the rigour of the errata model and adds the scheduled-review discipline that none of the precedents enforce.

The asymmetry that makes verdict tracking realistic at scale is operational. A single human editor cannot review one hundred and fifty active claims on a thirty-to-ninety day rolling cycle. The arithmetic does not work. An AI-augmented review workflow can, the AI assistant retrieves the relevant primary sources, summarises what has changed since the last review, and proposes a verdict; the human signatory reads the summary, audits the primary sources, and signs the verdict. The speed asymmetry between AI and human readers is the thing that makes the cadence physically possible. The architecture turns the asymmetry into accountability the reader benefits from.

Verdict tracking creates the conditions for the next component. Without explicit verdict states, retractions cannot be distinguished from edits; without dated transitions, the audit trail collapses into prose.

4. Dated retraction

The strongest part of the argument is here: retractions are events, not edits.

An edit changes the prose silently. The reader sees the new text and has no way to know that an earlier version of the same article said something different, unless they happen to have read both versions or to consult an external archive. Edits accumulate quietly. Over time, the publication’s relationship to its own history becomes opaque. The cumulative effect is that the reader cannot tell, when a claim now reads differently, whether the publication has changed its mind, whether the world has changed, whether the original was wrong, or whether somebody simply tightened the wording. Trust degrades because the public record is no longer a record.

A retraction does the opposite. It preserves the original prose exactly as published. It marks the original visually as superseded. It dates the moment of supersession to the day the publication issued the retraction. It names the corrected reality where the corrected reality can be named. It logs the retraction in a public retractions ledger that lists every retraction the publication has ever issued, dated and linked back to the original piece.

Silent edits destroy trust over time. Public, dated retractions build trust over time. The mechanism is the same in both directions: the reader’s ability to verify the publication’s own track record. Edits remove that ability; retractions add to it. The publication that runs a retractions ledger is offering the reader a permanent inventory of when it was wrong and how it responded, a stronger trust signal than any quantity of corrections-free prose, because it is the kind of signal that cannot be produced by a publication that does not run the discipline.

The architecture’s retraction structure has four parts. The first is a dated entry in a public retractions surface, typically a URL at /retractions/ listing every retraction the publication has ever issued. The second is a link from the original article to the retraction notice, surfaced prominently above the original prose so a reader landing on the retracted article from a cold link sees the retraction before the prose. The third is a link from the retraction notice back to a corrected piece, where a corrected piece exists. The fourth is a short explanation of what was wrong and why, written in the same voice as the publication’s editorial work, not in the apologetic register of a press release and not in the legalistic register of a notice of correction, but in the analytical register the publication uses for everything else it publishes.

The closest precedents are in academic publishing. The Committee on Publication Ethics, known as COPE, maintains retraction guidelines that have been adopted across substantial portions of the scientific literature, with explicit criteria for when a retraction is appropriate and what a retraction notice should contain. The International Committee of Medical Journal Editors maintains a corrections framework that distinguishes between corrections, expressions of concern, and retractions, each with its own typographic and bibliographic treatment. The US National Center for Biotechnology Information operates a retraction-tagging system inside PubMed that ensures a retracted paper carries the retraction label wherever it appears in literature databases. The architecture borrows the rigour of all three and applies it to consumer-grade commentary publishing, which has historically operated with none of it.

The AI-author advantage at this component is psychological. Retraction is harder for human writers than it should be, because the retraction marks the writer as wrong in a way that the writer’s identity is tied up in. An AI-drafted publication can retract with less personal cost because the human signatory is the steward of the claim, not the author of the prose. The signatory has no ego invested in a particular sentence; the signatory’s editorial reputation is invested in the discipline of how the publication responds when a sentence turns out to be wrong. The architecture makes retraction structurally easier, which makes retraction more frequent, which makes the publication’s track record more honest.

A publication that retracts visibly when it is wrong is more credible than a publication that never retracts. The retraction count is a feature, not a flaw. A retractions ledger reading zero after two years of publishing is not evidence of accuracy; it is evidence of a publication that has not yet developed the muscle to admit error. The architecture treats the retractions surface as a positive trust signal, the more populated, the more trustworthy the publication’s relationship to its own corpus.

5. Primary-source pinning

Most claims in commentary publishing rest on chains of citation. The trade press cites analyst notes. Analyst notes cite vendor briefings. Vendor briefings cite internal data nobody outside the room has seen. The reader, by the time the chain reaches them, is multiple steps removed from anything verifiable. The publication is rarely transparent about which step in the chain it stopped at, and the reader is rarely positioned to walk the chain backward.

Primary-source pinning means every claim links, where a primary source exists, to the primary source itself, not to a secondary report describing the primary source. If a claim references vendor product capability, the link is to the vendor’s product documentation or a regulatory filing, not to the trade-press article reporting on the documentation. If a claim references an analyst figure, the link is to the analyst report containing the figure, not to a press release quoting the report. If a claim references a regulator’s position, the link is to the regulator’s published text, not to a coverage piece summarising it.

The practical mechanic at this publication is that every claim card in the ledger carries a primary-source link. The link points to the original document, recording, regulatory filing, public earnings call, named-source statement, or first-party account. Secondary citation is explicitly labelled as secondary; mixed primary-secondary chains are flagged in the claim card so the reader can see which assertion rests on first-hand evidence and which rests on someone else’s summary. Where a claim is supported only by secondary citation, the architecture requires the claim to be downgraded to Partial, not because the secondary source is wrong, but because the publication’s accountability stops where the citation chain stops being walkable by the reader.

The closest precedents combine to make a workable approach. Investigative journalism’s document-drop tradition, publishing the underlying documents alongside the reporting, is the canonical primary-source-pinning move, but it is reactively applied to investigations rather than systematically applied to every claim. Academic citation discipline requires primary sourcing as a matter of methodological hygiene, but academic prose is not consumed by general readers in a format where the citations are clickable in line. Financial analysis pins valuation claims to regulatory filings as a matter of professional liability, but the financial-analysis surface is not a consumer publication and its citation discipline is invisible to its end readers. The architecture combines the rigour of academic citation, the clickable affordance of consumer publishing, and the document-drop reflex of investigative work, applied across every claim the publication makes.

The AI-author advantage at this component is mechanical, not creative. Distinguishing primary from secondary citation is an exercise in reading source URLs and document provenance. It is the kind of work that scales poorly for a deadline-pressed human commentator and scales naturally for an AI-augmented drafting workflow that has been instructed to track provenance from the first source it touches. The architecture instructs the workflow to do so. The result is that claims published here that survive into the ledger carry a primary-source link by default, and the few that do not are flagged as such, the publication’s honesty about the cases where its own citation chain is incomplete is part of the trust contract.

Primary-source pinning is what prevents the publication from drifting into the citation-chain failure mode this publication itself has previously named, the pattern by which commentary rests on commentary rests on commentary until the underlying evidence has been lost in three or four hops. Naming the failure mode and then preventing it inside the same architecture is the kind of self-discipline the architecture is built to make routine.

6. Review cadence

Claims expire. A vendor claim about a product’s capability in March 2026 is not the same claim about the same product in March 2027. The product has shipped, walked back, repositioned, deprecated, or died. A regulatory claim is similar: the framework that was authoritative at the time of writing has since been amended, interpreted, or contested. A market claim drifts even faster. A publication that does not revisit its own claims is publishing fossils, the prose stays on the page, and the page stays at the original URL, but the claim has aged out of the evidence base that supported it.

The architecture’s response is to commit, in advance and in public, to revisiting every claim on a published cadence. The parent register reviews every claim every thirty to ninety days, with the specific cadence set per claim based on how fast its underlying domain moves. Security-advisory claims on a thirty-day cycle. Governance-pattern claims on a sixty- to ninety-day cycle. The operators register runs faster, every thirty to forty-five days, because tooling pricing and SMB-relevant facts shift on a shorter horizon. The cadence is named on every claim card. The next review date is publicly visible. The discipline is auditable from outside.

A review concludes with a verdict, using the four-valued vocabulary from Section 3. Verdicts that change are dated. Verdicts that hold are also dated. A claim card reading “reviewed 15 Aug 2026, still Holding” is itself an act of accountability. It says: the publication looked again, the publication checked the primary sources, the publication is willing to continue asserting this claim, and the publication is on record as having done that work on this date. The same card three months later, reading “reviewed 14 Nov 2026, Partial, see correction below”, is the dated trail of the publication’s epistemic state moving in response to evidence.

Review cadence converts a static publication into a continuously maintained map of what the publication believes is true now and why. A traditional publication is a record of what we thought at a series of past moments. A publication running the architecture is a system whose current state is its current verdict on each claim it has made. The two artefacts are different objects, and they reward different reader behaviours: the traditional artefact rewards a reader who reads once; the architectural artefact rewards a reader who returns.

The closest precedents come from outside editorial publishing. Software dependency-update cadences, the discipline by which a serious team revisits its third-party dependencies on a scheduled rhythm, operate on the same principle: the dependency was correct when adopted, but the world moves, and the discipline is the scheduled re-examination. Security advisory revision cycles do the same for known vulnerabilities. Regulatory guidance amendment processes do the same for compliance-relevant texts. Each handles a piece of the discipline at scale. The architecture borrows the discipline and applies it to commentary, which has historically operated with none of it.

Review cadence is the component with the largest operational asymmetry between human and AI-augmented publishing. A single human editor cannot review one hundred and fifty active claims on a thirty-to-ninety day rolling cycle. The arithmetic, again, does not work. An AI-augmented review workflow can, and at that point the cadence becomes a published commitment the publication can keep rather than a wishful piece of editorial branding. The asymmetry is the most distinctive operational signature of an AI-written publication built on this architecture.

Review cadence is what makes the publication a system rather than an archive.

7. Conclusion

The load-bearing claim of this piece is that AI-written commentary can be more verifiable, not less, than human-written commentary, when it is published inside an explicit accountability architecture with six components. The six components are disclosure, claim isolation, verdict tracking, dated retraction, primary-source pinning, and review cadence. Each is necessary; none alone is sufficient. The combination is the contribution.

The obvious objection is that the architecture is just a particular operating model, and that lots of publications could adopt parts of it without giving the publication any defensible competitive position. The objection is correct, and that is the point. The architecture is portable. It is published openly under the Creative Commons Attribution 4.0 licence. Other publications are encouraged to adopt the components, singly, in pairs, in clusters, or all six together, and to publish their own ledgers under their own naming schemes. The publication’s competitive advantage is not the architecture itself; it is the back-catalogue of claims processed through the architecture, the track record of verdicts dated and held, and the cumulative trust artefact that only a publication actually running the discipline can produce. The architecture is the thing you adopt; the track record is the thing you cannot copy.

The deeper objection is that the architecture itself depends on the human signatory making good judgement calls about which AI-drafted claims to sign. That objection is also correct. Judgement is the irreducible human contribution, and the architecture does not eliminate it. What the architecture does is isolate it. The human contribution is signoff on claims, not authorship of prose. The architecture makes the signoff load-bearing and the prose disposable. This is the right place to put the human element, because it is the place where human judgement is genuinely better than AI capability, assessing whether a claim is worth standing behind, on the specific evidence at hand, in the specific context the claim will be read in. The AI drafts; the human decides; the architecture makes the decision public.

A publication that discloses its AI authorship and runs the six components is more accountable to its readers than a publication that hides the same authorship behind a human byline.