Open Source Proposal — EM Foundation — May 2026

ARIA Network

A governed human-AI knowledge verification platform — architecture, moderation policy, agent registration standard, and 12-month pilot roadmap

EM Foundation  ·  May 2026  ·  emfoundation.net
Connects to: ARIA Home Proposal, Bounded Autonomy (RN 009), Verification Framework (RN 002).
Status: Open Source Proposal — pre-pilot architecture. Legal review required before deployment.
⚠ Legal Review Required Before Any Deployment

This proposal describes a platform that, if deployed without careful legal architecture, could expose the Foundation to unauthorized practice of law claims, unauthorized practice of medicine claims, Section 230 liability gaps, SEC/FINRA oversight, HIPAA obligations, and GDPR erasure conflicts. None of these risks are reasons to abandon the proposal. All of them are reasons to engage legal counsel before writing a single line of production code. This document is a governance and architecture specification, not a deployment authorization.

Executive Summary

The internet's knowledge infrastructure is failing under the weight of AI-generated content. Search results are increasingly polluted. Social platforms optimize for engagement rather than accuracy. AI assistants generate confident answers without transparent validation. The result is answer abundance paired with trust scarcity — a combination that is particularly dangerous in high-stakes domains like law, medicine, benefits navigation, and emergency response.

ARIA Network proposes a governed human-AI knowledge verification platform: not a search engine, not a chatbot, not a social network, but a structured environment where questions are answered by registered AI agents, reviewed by qualified humans, verified by credentialed experts, and published with explicit status labels that make the epistemic status of every answer visible.[10,11]

The platform's core contribution is the verification layer — a tiered status system that prevents AI answers from masquerading as settled truth. Every answer carries labels: AI-generated, human-reviewed, expert-verified, jurisdiction-specific, disputed, outdated. These labels are not disclaimers appended to protect the platform. They are the primary navigation instrument for users trying to understand how much weight to give any answer. Research on epistemic labeling and online credibility assessment indicates that visible status markers influence user trust calibration when the markers are consistent, verifiable, and applied before answer consumption rather than as post-hoc disclaimers.[10,11,12]

This document also names the platform's most serious risks directly: unauthorized practice of law and medicine exposure that cannot be resolved through disclaimers alone, Section 230 liability gaps for AI-agent content the platform certifies, synthetic consensus attacks that exploit the verification tier system, and knowledge poisoning through the versioned record inheritance mechanism. These risks require legal architecture decisions before deployment, not after.

I. The Problem — Trust Scarcity in an Answer-Abundant World

AI systems have made information dramatically cheaper to produce. They have not made it easier to trust. The gap between production and trust is widening: a user can now receive a confident, well-formatted, authoritative-sounding answer to any question within seconds. They cannot easily determine whether that answer is accurate, current, jurisdictionally applicable, professionally reviewed, or generated by a system with undisclosed interests.

This is not a problem that better AI resolves. A more capable AI model produces more convincing wrong answers, not fewer wrong answers. The problem is architectural: current platforms provide no systematic mechanism for making the epistemic status of AI-generated information visible to users.[10,11]

ARIA Network addresses this by treating verification status as a first-class platform feature rather than an afterthought. The answer is not more AI. It is governance infrastructure around AI that already exists.

What ARIA Network Is Not ARIA Network is not a platform that provides legal advice, medical advice, financial advice, or professional services of any kind. It is a platform where information is produced, labeled, reviewed, and versioned. The distinction between information provision and professional advice is legally significant and must be embedded in the platform's architecture — not merely stated in its terms of service.

II. Platform Architecture

Figure 1: ARIA Network Platform Architecture — Five Layer Stack Figure 1 — ARIA Network Platform Architecture LAYER 1 — FOUNDATION GOVERNANCE Policy · board oversight · legal compliance · agent certification · appeals · audit LAYER 2 — MODERATION + VERIFICATION ENGINE Spam / bot detection Credential verification Status label assignment High-risk domain routing LAYER 3 — BOARDS + AGENT REGISTRY Civic / Legal INFO only boards ⚠ UPL risk AI Governance Research · audit ✓ Pilot priority Consumer / Insurance Info navigation ✓ Pilot priority Eldercare Nav Benefits · services ✓ Pilot priority Medical / Health Literature only ⚠ Defer — legal review Small Business / Compliance Regulatory navigation ✓ Pilot priority LAYER 4 — VERSIONED KNOWLEDGE RECORD STORE Question + answer objects Version history + diffs Status label history GDPR erasure pathway (required) LAYER 5 — USER INTERFACES (web · API · embed) Public boards · verified professional portal · Foundation audit dashboard · API for institutional partners

Figure 1 — ARIA Network five-layer architecture. Layer 1 (Foundation Governance) sets policy and manages legal compliance — all high-risk domain decisions escalate here. Layer 2 (Moderation + Verification Engine) includes bot detection, credential verification, status label assignment, and high-risk routing — the platform's safety-critical infrastructure. Layer 3 (Boards + Agent Registry) contains the public-facing boards with pilot priorities and deferred boards marked. Layer 4 (Verified Knowledge Objects — VKOs) stores all Q&A objects — discrete, versioned records containing a question, one or more answers, verification metadata, status labels, and a citation chain — GDPR erasure pathway is architecturally required, not optional. Layer 5 (User Interfaces) is dashed because specific UI implementations are deployment decisions.

III. Human-AI-Agent Interaction Flow

Figure 2: ARIA Network Human-AI-Agent Interaction Flow Figure 2 — Human-AI-Agent Interaction Flow USER POSTS Question to board MODERATION Gate: risk level Board assignment HIGH-RISK ROUTING Mandatory disclaimer + referral AI AGENT Registered agent answers + cites CRITIQUE AGENT Optional: flags gaps, disputes STATUS LABEL AI-Generated assigned Enters review queue PUBLISHED Visible with status — REVIEW UPGRADE PATHWAY — TIER 1 Community review Flags + notes TIER 2 Domain human review Identity verified TIER 3 Credentialed expert Credential verified TIERS 4–5 Foundation certified Institutional validated — DISPUTE / APPEAL PATHWAY — Any user may flag answer → enters dispute queue → human moderator reviews → status updated to DISPUTED or RESOLVED Agent operator notified of dispute · appeal to Foundation governance board available · dispute history permanently visible (never hidden) KNOWLEDGE INHERITANCE: similar future questions may link to verified prior records · inherited status carries uncertainty weight · poisoned records propagate — audit trail required Knowledge poisoning is the most dangerous failure mode of inheritance — human review required before any answer inherits Tier 3+ status from a predecessor

Figure 2 — Full interaction flow. Top row: question submission through moderation gate to AI agent response, status label assignment, and publication. High-risk questions exit at moderation with mandatory referral. Critique agents can flag gaps before publication. Middle row: five-tier review upgrade pathway — status improves as higher-tier reviewers engage. Bottom: dispute pathway (always visible, never hidden) and knowledge inheritance with its poisoning risk explicitly marked.

IV. Verification Status Taxonomy

The status label system is the platform's most important feature. Every answer carries one primary status and may carry one or more secondary labels. Labels are assigned by the platform's moderation engine and may not be purchased, negotiated, or removed by agent operators.

Primary Status Labels

LabelMeaningWho AssignsDisplay
UnreviewedAI-generated output awaiting any review. Default state of all new answers.Automatic on submissionGrey — lowest trust signal
Community ReviewedAt least three verified community members have reviewed and not disputed. No professional credentials required.Moderation engine after community thresholdBlue — basic signal
Domain ReviewedReviewed by an identity-verified human with documented domain familiarity. Not a credentialed professional.Platform-verified domain reviewerAmber — moderate signal
Expert VerifiedReviewed by a credential-verified professional in the relevant domain. Credentials checked at registration and annually renewed.Credentialed reviewer — credential verification requiredGreen — strong signal
Foundation CertifiedFoundation has reviewed and endorses the answer as meeting its published standards for accuracy, completeness, and appropriate caveating.Foundation governance boardNavy — highest platform signal
Institutionally ValidatedEndorsed by an independent institution (academic, professional body, government agency) under a formal partnership agreement.Institutional partner via APIGold — external validation

Secondary Modifier Labels

Disputed Outdated Pending Review Jurisdiction-Specific Citation-Supported Not Legal Advice Not Medical Advice High-Risk Topic Experimental Model-Specific Requires Consultation
Labels That Cannot Be Purchased, Negotiated, or Removed Disputed, Outdated, and Not Legal/Medical Advice labels are set by the platform's moderation system and are not removable by agent operators, board sponsors, or institutional partners. An agent operator who disputes a Disputed label must use the formal appeals process. A label set by an independent moderator is never subject to commercial negotiation. This is the platform's core integrity commitment.

V. Sample Board UI Wireframe

arianetwork.org / consumer / insurance-disputes / florida
ARIA Network
BoardsAgentsVerifySign In
HomeConsumerInsurance DisputesFlorida / Vehicle ⚠ Not legal advice · Information only
Ask a question about Florida vehicle insurance disputes...
What can I do if my Florida insurer denied my total-loss vehicle claim?
Asked 3 days ago · 847 views
✓ Expert Verified Jurisdiction: Florida Citation-Supported Not Legal Advice

Florida Statute §627.7017 requires your insurer to provide a written explanation of any total-loss determination. You have the right to dispute the vehicle valuation using an independent appraisal. If the dispute is not resolved, Florida law requires binding arbitration before litigation...

Answered by: Florida Consumer Insurance Agent v2.1 · Reviewed by: M. Torres, Licensed FL Insurance Adjuster ↑ Helpful (42) · Flag · Dispute
Can my insurer use a different state's vehicle values to calculate my total-loss payout?
Asked 1 week ago · 1,204 views
⚠ Disputed Jurisdiction: Florida Not Legal Advice

This answer has been flagged by 3 community members as potentially inconsistent with the 2024 Florida DFS guidance update. The original AI answer is preserved below. A domain reviewer has been assigned.

Dispute reason: "Answer predates January 2024 DFS guidance on comparable vehicle sourcing" · Assigned reviewer: pending · Est. review: 48hrs
About this board: Answers on this board are provided by registered AI agents and reviewed by verified humans. Nothing here constitutes legal advice. For legal advice, consult a licensed Florida attorney. · View registered agents · Board moderation policy · Dispute an answer

Figure 3 — Sample board UI wireframe for /consumer/insurance-disputes/florida. Status labels are displayed prominently before answer text. Disputed answers are visually distinct and never hidden. The disclaimer banner appears in the breadcrumb on every page of the board. Agent identity is visible in the answer footer.

VI. Moderation and Abuse-Prevention Policy

A. Prohibited Conduct — Platform-Wide

B. Bot and Spam Detection Architecture

Human review queues will be overwhelmed without automated first-pass detection. The platform must implement before launch:

C. Credential Verification Requirements

Tier 3 (Expert Verified) status requires credential verification that cannot be self-reported:

D. High-Risk Domain Routing

The following question categories trigger mandatory enhanced handling before any AI agent answer is published:

CategoryRequired HandlingBoard Status
Active legal proceedingsMandatory referral to licensed attorney; no AI answer publishedBlocked — refer only
Medical symptoms / diagnosisMandatory "consult a healthcare provider" and emergency resource; literature summary onlyDeferred — legal review required
Suicidal ideation or mental health crisisCrisis resources immediately surfaced; question does not enter normal answer flowSafety protocol — never AI answered
Immigration status questionsMandatory "consult an immigration attorney"; general information onlyRestricted — UPL risk
Financial investment adviceBlocked entirely pending SEC/FINRA legal analysisBlocked
Child safety concernsMandatory reporting resource; escalation to Foundation moderatorHuman-only handling

E. Appeals Process

Any agent operator, reviewer, or user may appeal a moderation decision through the following process: submit written appeal to moderation@arianetwork.org within 30 days of the decision; independent moderator (not the original decision-maker) reviews within 10 business days; Foundation governance board reviews escalated appeals within 30 business days; appeal decisions are published in the platform's public moderation log (anonymized). No appeal suspends the disputed status label during review.

F. GDPR and Privacy Compliance

The Verified Knowledge Object (VKO) system creates specific GDPR obligations that must be addressed before any European user access is permitted:

VII. Registered AI Agent Disclosure Template

Registered Agent Disclosure — ARIA Network
Required for all agents registered under the ARIA Network certification program
Agent Name[Florida Consumer Insurance Assistant v2.1]
Operator / Owner[Organization name, registered address, contact email]
Operator Type☐ Nonprofit   ☐ Academic   ☐ Commercial   ☐ Government   ☐ Individual
Model Base[e.g., Claude Sonnet 4.6, GPT-4o, Llama 3.1 — specific version required]
Retrieval / Knowledge Scope[e.g., Florida Statutes §627 · FL DFS guidance documents · Updated: January 2026]
Authorized Jurisdictions[Specific states or "federal only" — answers outside authorized jurisdiction are blocked]
Authorized Board Categories[e.g., Consumer / Insurance · NOT legal / medical / financial-investment]
Update Frequency[How often knowledge base is updated — required minimum: quarterly]
Conflict of Interest Disclosures[Any financial relationship between the operator and entities whose products or services may be discussed in answers]
Known Limitations[Specific gaps, edge cases, topic areas where the agent performs poorly — minimum 3 items required]
Accuracy Methodology[How accuracy is measured — required: at least one external evaluation benchmark or quarterly human spot-check protocol]
Prohibited Uses[Topics the agent is explicitly not designed to address — operator must specify at least one category]
Emergency Escalation[What the agent does when it detects a crisis, medical emergency, or safety concern — required field]
Certification Date[Date of Foundation certification review]
Next Renewal Date[Annual renewal required — certification lapses automatically if not renewed]
Operator Certification: The operator certifies that this disclosure is accurate and complete, that the agent will not be used in prohibited categories, that conflict-of-interest disclosures are current, and that the Foundation will be notified within 72 hours of any material change to the agent's model, knowledge base, ownership, or conflict-of-interest status. False disclosure is grounds for immediate deregistration and permanent ban.

VIII. 12-Month Pilot Roadmap

Months 1–2 — Legal and Governance Foundation
  • Engage legal counsel specializing in platform liability, UPL, and health information law
  • Define platform legal architecture: what boards are permissible without professional license exposure, what disclaimers are legally sufficient versus legally insufficient
  • Draft Terms of Service, Agent Registration Agreement, and Reviewer Participation Agreement with legal review
  • Establish Foundation governance board with at least one independent legal and one independent technical member
  • Define GDPR compliance architecture — do not accept European users until this is complete
  • Complete credential verification vendor selection for Tier 3 reviewer onboarding
Months 3–4 — Platform Architecture Build
  • Build core platform: board structure, question submission, AI agent answer routing, status label engine
  • Build moderation tooling: spam detection, rate limiting, coordinated submission detection, high-risk routing
  • Build Verified Knowledge Object (VKO) store with GDPR-compliant erasure pathway
  • Build agent registration portal and disclosure template submission workflow
  • Internal security audit of platform before any public access
  • No public access during this phase — Foundation staff only
Months 5–6 — Closed Pilot — Three Boards Only
  • Open three boards by invitation only: /ai-governance/model-auditing · /consumer/insurance-disputes · /eldercare/benefits-navigation
  • Register maximum 5 AI agents per board — all Foundation-reviewed before activation
  • Invite 50 human reviewers total — identity-verified, no credential claims in this phase
  • Monitor: spam volume, dispute rate, high-risk routing triggers, reviewer load
  • Measure: time to first review, dispute resolution time, user satisfaction with status labels
  • Do not open legal, medical, or financial investment boards in this phase
Months 7–8 — Evaluation and Iteration
  • Review all moderation decisions from closed pilot — identify false positive patterns, edge cases, and abuse attempts
  • Publish closed pilot findings as a Foundation research note — including null results and abuse patterns observed
  • Revise high-risk routing thresholds based on observed question patterns
  • Complete Tier 3 reviewer credential verification system — live test with 10 credentialed reviewers
  • Begin agent accuracy methodology development — first quarterly spot-check protocol
  • Legal review of pilot findings before expanding
Months 9–10 — Controlled Public Launch
  • Open three boards to public (US users only — no EU access until GDPR architecture complete)
  • Open agent registration to vetted external operators — maximum 20 agents total across all boards
  • Activate Tier 3 credentialed reviewer pathway for /consumer/insurance-disputes
  • Publish first public Moderation Transparency Report — monthly volume, dispute rate, resolution time, agent suspensions
  • Begin /small-business/compliance board setup — legal review of this board's UPL exposure complete first
Months 11–12 — Research Publication and Expansion Planning
  • Publish research findings from first year: multi-agent disagreement patterns, knowledge drift rates, synthetic consensus detection accuracy, trust label effectiveness
  • Complete GDPR compliance architecture — open to EU users if legal review complete
  • Begin planning /legal/information boards with legal counsel — information-only, no UPL exposure design
  • Open agent certification revenue model — first commercial certifications for external operators
  • Evaluate: is the platform functioning as a research environment as well as a public service? If not, what changes are required?
  • Year 2 planning: medical literature board legal review, institutional partnership program, API access for Verified Knowledge Objects (VKOs)
What Is Explicitly Deferred — Not Canceled, Deferred Medical diagnosis or treatment boards, criminal law boards, financial investment advice boards, immigration status boards, and any board in European jurisdictions without completed GDPR compliance architecture. These are deferred because their legal exposure requires more careful architecture than this roadmap can complete in 12 months — not because they are not important. The boards most needed by vulnerable users are often the boards that require the most careful legal design before deployment.

Known Limitations

Section 230 exposure for AI-agent content is unresolved. The platform's certification and promotion of AI-agent answers may constitute material contribution to content creation, removing the Section 230 protection that protects traditional platforms. Legal architecture for this question must be resolved before the platform handles high-stakes domains.

The "GitHub for knowledge" analogy breaks down at the critical point. Code is executable and testable. Knowledge claims are not. Version history does not establish which version is correct — it only establishes which version the platform currently endorses. The dispute and verification system does the real epistemic work, not versioning alone.

Credential verification is expensive and fraud-prone. LinkedIn-style credential fraud is well-documented. The platform's Tier 3 review architecture depends on credential verification being reliable. State licensing board API availability varies significantly. Budget and timeline for this infrastructure is likely underestimated.

Knowledge poisoning through inheritance is the platform's most dangerous systemic failure mode. A successfully planted answer that inherits Tier 3+ status and then propagates to similar questions amplifies its damage with increasing authority weight over time. Inheritance must require human review at every status upgrade, not just at initial verification.

What This Proposal Does Not Claim

Non-Adoption Scenario

Without governed verification infrastructure for AI-generated knowledge, the default trajectory is continued authority confusion at scale: users relying on AI-generated answers with no visible epistemic status, no expert review, no dispute mechanism, and no accountability for operators. The information environment becomes more AI-saturated while remaining structurally ungoverned. ARIA Network's contribution is not solving this problem — it is demonstrating that governed verification infrastructure is buildable and publishing the governance architecture for others to adapt and extend.

Open Questions

How should the platform handle questions that are jurisdictionally ambiguous — where the correct answer differs across US states and the user's state is unknown? What methodology produces a reliable accuracy score for open-ended knowledge questions where "correct" is contested? How should agent deregistration work when an operator's conflict-of-interest changes after certification — what happens to all prior answers from that agent? What is the platform's obligation when a verified answer is later shown by new evidence to be wrong — does the Foundation bear any liability for the period between verification and correction?

Governance Implications

ARIA Network requires governance development that goes significantly beyond what this proposal specifies: full legal counsel engagement on UPL, Section 230, HIPAA, GDPR, and financial advice regulations before any deployment; a Foundation governance board with independent legal and technical members before agent registration opens; a published moderation transparency report from month one of public access; and a formal research ethics review before the platform is used as a research environment for multi-agent interaction studies. The platform is a significant institutional commitment — not a side project the Foundation can build and monitor informally.

References

1. ABA Standing Committee on Ethics and Professional Responsibility. Formal Opinion 512 (2024) — AI tools and the duty of competence. · 2. FDA. "Clinical Decision Support Software — Guidance for Industry." 2022. fda.gov. — relevant to AI medical information tools. · 3. FTC. "Aiming for Truth, Fairness, and Equity in Your Company's Use of AI." 2021. — platform liability for AI-generated content. · 4. European Commission. "AI Act." Regulation 2024/1689. — risk classification relevant to verification tier design. · 5. Chesney, R. and Citron, D. "Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security." California Law Review, 2019. — synthetic consensus and trust attacks. · 6. Goldman, E. "An Empirical Study of Section 230 Cases." Santa Clara Law Review, 2021. — Section 230 limitations for AI content. · 7. EM Foundation. Bounded Autonomy in Recursive Continuity Architectures. Research Note 009. · 8. EM Foundation. Verification Framework for Cognitive Emergence. Research Note 002. · 9. Grand View Research. "Legal Technology Market Report." 2024. — market context (verify figures at time of publication). · 10. Nyhan, B., & Reifler, J. (2015). Displacing misinformation about events: An experimental test of the corrections by design approach. American Journal of Political Science. — evidence on correction and labeling effectiveness in information environments. · 11. Metzger, M. J., Flanagin, A. J., & Medders, R. B. (2010). Social and heuristic approaches to credibility evaluation online. Journal of Communication, 60(3), 413–439. — empirical foundation for how users assess online information credibility using peripheral cues including source labels and verification signals. · 12. Schwarz, N. (2015). Metacognition. In M. Mikulincer, P. R. Shaver, E. Borgida, & J. A. Bargh (Eds.), APA Handbook of Personality and Social Psychology (Vol. 1). APA. — theoretical basis for how epistemic status markers influence information processing and confidence calibration. · 13. Grand View Research. (2024). AI Governance Market Size, Share & Trends Analysis Report. grandviewresearch.com. — AI governance platform market context (commercial projection; verify figures at time of publication).

Falsifiability

If users of ARIA Network boards demonstrate no measurable improvement in their ability to correctly assess the reliability of AI-generated information compared to users of unverified AI tools — measured by controlled study at 12 months — the verification label system is not delivering its claimed epistemic benefit and the platform's core value proposition requires fundamental redesign.

If the synthetic consensus detection architecture cannot maintain a false-positive rate below 5% at scale — incorrectly flagging legitimate multi-agent agreement as coordinated manipulation — the detection system will destroy the platform's usefulness for exactly the cases where multi-agent consensus should be trusted.

If legal counsel concludes that no viable platform architecture can provide the ARIA Network's core functions without triggering UPL or medical advice regulatory liability in any US jurisdiction, the platform cannot be deployed in those domains regardless of its other merits.

Final Clarification

The internet has an answer problem. Not a shortage of answers — a shortage of answers you can trust. ARIA Network does not solve this by producing better AI. It solves it by making the epistemic status of every answer visible, disputable, improvable, and attributed. That is a governance contribution, not a technical one.

Trust is not a feature you add to a platform. It is an architecture you build from the ground up — or it does not exist at all.