How ARX Intelligence is written, audited, and corrected.

ARX Intelligence is a feed of short commentary on the AI infrastructure, compliance, agent-coordination, and enterprise-AI market landscape. The commentary text on this site is generated by a large language model (Claude, from Anthropic), grounded in a retrieval pipeline of public sources, and reviewed by additional models and by ARX staff before publication.

It is not a wire service, not legal advice, not investment advice, and not a substitute for primary reporting. The original articles we comment on are linked from every item; please read them at the source.

ARX QM Holdings, Inc. is the publisher and is responsible for the editorial decisions made about what runs and what does not run. Editorial control rests with ARX staff, not with the models that draft and review the text.

Every published item moves through the same pipeline:

1. Sourcing. An automated process pulls article candidates from public feeds and filters them for relevance to the topics we cover.
2. Grounding. A retrieval step gathers public-source context that the model can cite. The model is instructed to only make claims it can ground in that retrieved context or in widely established public knowledge.
3. Drafting. A writer model (Claude Sonnet) drafts the commentary and a list of citations.
4. Claim audit. A second model decomposes the draft into individual claims and rates each as supported, speculative, or unsupported, producing a numeric risk score.
5. Adversarial review. A third, more capable model (Claude Opus) red-teams the draft for defamation risk, hallucinated quotes or numbers, brand-voice violations, stale or invented build claims about ARX, copyright concerns, and securities-adjacent statements.
6. Tier classification. Audit and adversarial outputs are combined into a tier: GREEN (low risk), YELLOW (review-before-publish), or RED (manual edit required). Audit or adversarial failure defaults to RED.
7. Editorial decision. GREEN items may auto-publish. YELLOW items are reviewed by ARX staff before publication. RED items are manually edited or held. Every editorial action is logged.

Adversarial review checks each draft for the following risks explicitly, not as a general “is this safe” prompt:

• Defamation. Statements about identifiable people or companies that assert facts presented as established, when those facts are not supported by the retrieved sources, are flagged.
• Hallucination. Direct quotes, specific dollar figures, dates, percentages, and named-source attributions are checked against the retrieved sources.
• Brand-voice violations. Commentary that misrepresents ARX's positioning or attributes positions to ARX that ARX has not taken is flagged.
• Stale or invented build status. Claims about what ARX has shipped, plans to ship, or has proven are checked against the canonical project status.
• Copyright. Long verbatim excerpts from source articles are flagged. Commentary should add analysis, not reproduce source text.
• Securities-adjacent statements. Statements that could be read as investment advice, price predictions, or material non-public information about publicly traded companies are flagged.

Every commentary item carries an inline AI disclosure. The disclosure text is versioned. When the disclosure changes, in-flight commentary is re-tagged so that any later correction or recall is issued against the right baseline.

When ARX learns that a published item contains a factual error, an attribution error, missing context, or other material problem, we issue a correction. Corrections are attached to the original item, are also listed at /intelligence/corrections, and remain attached for the life of the item.

If a correction is severe enough that the item should no longer be presented as ARX commentary at all, it is recalled.

A recalled item is removed from the public feed, returns HTTP 410 Gone for direct requests (so search engines drop it), and the original URL displays a public recall notice explaining why the item was recalled. Recalls are recorded with the same audit trail as publication.

The pipeline carries automated kill switches. If the moderation queue depth, the rejection rate, or the RED-tier rate exceeds preset thresholds, the pipeline halts on its own and stops producing new commentary until a human clears the kill state. These thresholds exist so that a model regression or a source-feed problem cannot silently fill the site with low-quality output.

For each published item we retain: the source article, the model draft, the audit decomposition, the adversarial review, the tier decision, the editor (if any) and editor action, the disclosure version in effect at publication, and any corrections or recalls. This record is the basis for every claim we make about how this system runs.

We do not sell, rent, or share this editorial record outside of ARX, except as required by law or in response to a legitimate legal process.

To request a correction, a retraction, or to report a problem with a specific item, contact us via the /contact page. Please include the URL of the affected item and a description of the problem. We aim to respond to correction requests within two business days.

For legal notices, including DMCA notices and rights-holder inquiries, contact us through the same page and indicate that the message is a legal notice; it will be routed accordingly.

Material changes to this policy increment the policy version shown at the top of this page and update the effective date. The previous policy text is preserved internally and can be produced on request for any item published under it.