Most AI companies demo their product behind a sales call. Today we made ours public.
arxqm.com/intelligence is a live, AI-generated commentary feed covering enterprise AI, defense technology, and financial services governance. It updates automatically as significant developments break. It is free and open to anyone.
Every piece of commentary on that page was produced by a governed AI pipeline with a multi-stage editorial process, automated safety interlocks, and a complete compliance audit trail. The feed is not a marketing feature. It is the same infrastructure we are building for institutions that will be held accountable for what their AI systems produce, running in production, on content we are accountable for, visible to anyone who wants to evaluate it.
We want to explain why we built it this way, because the engineering decisions underneath this feed are a compressed demonstration of what ARX actually does.
The Problem the Feed Solves for Us
There is a credibility problem in AI infrastructure that we have been thinking about since before incorporation. Every vendor in this space claims governance. Every vendor claims auditability. Every vendor claims safety. Almost none of them can show you a live system producing real output under real accountability conditions, because the moment you do that, every failure mode you claim to handle becomes testable by anyone with a browser.
That is exactly the condition we wanted to create.
The intelligence feed puts our governance pipeline in a position where it has to work, publicly, continuously, on content that carries our name. If the AI hallucinates a claim, that claim appears on our domain. If the safety pipeline misses a risk, we own the consequences. If a commentary needs to be recalled or corrected, the recall and correction are themselves governed actions with audit trails. There is no staging environment for accountability. The production environment is the test.
We have watched enough enterprise software companies ship governance dashboards that look convincing in a demo and collapse under operational load. We did not want ARX to become one of them. The intelligence feed is our answer to the question every enterprise buyer should be asking and almost none of them do: show me a system you are willing to be held accountable for, running in production, right now.
What the Pipeline Does Before a Word Reaches You
The pipeline is not a single model generating text. It is a multi-stage process where separate AI systems perform distinct functions, each constrained by the output of the previous stage, with automated circuit breakers that can halt the entire pipeline if any safety threshold is breached.
Source retrieval. A hybrid engine pulls from 5 independent search channels, including semantic search, category search, and keyword search across the live web, fused with reciprocal rank fusion and passed through a neural reranker. The commentary is grounded in retrieved sources, not generated from training data. Every factual claim in the final output is traceable to a specific retrieved document.
Writing. The AI writes a first draft of commentary grounded exclusively in the retrieved sources and ARX’s internal knowledge base. The draft is opinionated. It takes a position on what the development means for the enterprise AI infrastructure landscape. That is intentional. A feed that summarizes headlines adds no value. A feed that interprets developments through a specific, stated lens and puts its analysis on the record is useful.
Claim auditing. A separate AI instance audits every factual claim in the draft, assigns a risk score, and flags anything that cannot be traced to a retrieved source or the knowledge base. This is not a grammar check. It is a claim-level verification pass that produces a structured risk assessment for the entire piece.
Adversarial review. A third AI instance reads the draft with the explicit instruction to find problems. It evaluates 6 specific risk categories: factual errors, unsupported claims, legal exposure, reputational risk, competitive mischaracterization, and bias. Any critical finding blocks publication. The adversarial reviewer does not know or care what the writer intended. Its job is to attack the draft, and its output determines whether the draft proceeds.
Tier classification. Based on the audit and review signals, each commentary is classified into one of 3 tiers:
- GREEN: Low risk, auto-publish. The audit passed, the adversarial review found no critical issues, and the confidence score is above threshold.
- YELLOW: Moderate risk, fast moderator review. A human must approve before publication, but the review is lightweight because the automated checks passed with caveats.
- RED: High risk, manual hold. A human must review the full audit trail and explicitly approve before publication.
Only GREEN items reach the public feed without human intervention. The tier system means the pipeline self-regulates its own output quality and routes anything uncertain to a human.
The Safety Interlocks
The tier classification handles individual commentary quality. The safety interlocks handle systemic failure.
If the pipeline’s queue depth exceeds a threshold, meaning more articles are arriving than the pipeline can process safely, the entire pipeline halts automatically. If the rejection rate crosses a threshold, meaning the adversarial reviewer is flagging too many drafts, the pipeline halts. If the proportion of RED-tier classifications exceeds a threshold, meaning the pipeline is producing too much high-risk content, the pipeline halts.
A human moderator can also halt the pipeline manually at any time. When triggered from the moderator interface, the halt propagates in under one second. The pipeline does not complete its current batch. It stops.
Every halt, whether automated or manual, is a governed action. It emits a compliance event that records who or what triggered the halt, why, what the threshold values were at the time, and when the halt was cleared. The moderator’s stated reason for a manual halt is persisted in the audit trail. Nothing is ephemeral.
The Audit Trail
This is the part that matters most, and the part that is hardest to explain without making it sound like a compliance checkbox exercise. It is not.
Every editorial action on the intelligence feed, approval, rejection, edit, recall, correction, kill-switch activation, kill-switch clearance, emits a CloudEvents v1.0 envelope to our compliance risk engine. The envelope is a structured, machine-readable record of what happened, who did it, when, and what the system state was at the time. The envelopes are append-only. They cannot be edited after the fact. They cannot be deleted.
This means we can reconstruct the complete editorial history of any piece of commentary on the feed: what sources were retrieved, what the AI wrote, what the auditor flagged, what the adversarial reviewer found, what tier it was classified into, who approved it, and if it was later recalled or corrected, why, by whom, and what the public notice said.
That capability is not something we built for the intelligence feed specifically. It is the same audit infrastructure we are building for enterprises operating under the Treasury’s Financial Services AI Risk Management Framework, where 230 control objectives require exactly this kind of traceable, reconstructable decision history. The intelligence feed is the first public deployment of that infrastructure. It will not be the last.
What Happens When Something Goes Wrong
We want to be explicit about this, because the failure handling is the architecture.
If a commentary on the feed contains an error, a moderator can recall it. The recall is a governed action with a structured reason (factual error, editorial judgment, legal or compliance concern, source reliability, market sensitivity, bias, technical error, or other). The recalled commentary is removed from the public feed with a notice that it was recalled. The recall action, the reason, and the moderator who initiated it are recorded in the audit trail.
If a commentary requires a correction rather than a full recall, a moderator can issue a correction. Corrections are typed (factual, editorial, source attribution, data or statistical, or legal or compliance). Each correction requires a public note of at least 10 characters explaining what was corrected. The original text and the corrected text are both preserved. The correction action, the type, the public note, and the moderator who issued it are recorded in the audit trail.
If a moderator issues a recall on a commentary that has already been corrected, or a correction on a commentary that has been recalled, the system rejects the action with an explicit conflict response. These are not edge cases we handle gracefully. They are illegal states the system makes impossible.
Why This Matters Beyond Our Feed
The intelligence feed is interesting on its own terms. It provides useful commentary on a domain we care about and that our market cares about. But the feed is not the point. The point is what the feed proves about the infrastructure underneath it.
Every enterprise deploying AI in a regulated environment will face the same set of requirements we imposed on ourselves: traceable sourcing, claim-level verification, adversarial review, tiered classification, safety interlocks, governed recalls and corrections, and an append-only audit trail that can satisfy an examiner, a regulator, or a litigator asking for the complete decision history of a piece of AI-generated output.
The Treasury’s FS AI RMF requires it. The EU AI Act’s Article 14 human oversight requirements assume it. CMMC and NIST 800-53 demand the kind of audit trail that makes every control enforceable at runtime rather than aspirational at the policy level.
Most organizations do not have this infrastructure. They know they will need it. They have not yet encountered a vendor who can demonstrate it in production, under real conditions, on output the vendor is willing to stake their name on.
We can. It is running right now. You can read it at arxqm.com/intelligence.
Previous in the series:
- The Stateful Runtime Layer for Enterprise AI
- The Infrastructure Is Converging
- ARX Hires Its First Chief Science Officer
- Treasury Just Told Every Bank in America to Govern Their AI
- The Architecture Cannot Keep the Promise
ARX is building the stateful runtime layer for enterprise AI. Governance, institutional memory, and cognitive portability across providers, models, and regulatory jurisdictions. Learn more at arxqm.com.
#AIGovernance #EnterpriseAI #DefenseAI #Compliance #FinancialServices #AIInfrastructure #ARX