Proposed Institutional Architecture

CODE-AI

Coalition of Domain Experts for Artificial Intelligence

Constitutional Architecture, Institutional Design and Research Program for Domain-Governed Artificial Intelligence

Author

Dr. Sharad Maheshwari MD

Think Tank

BeResponsibleAI & Institute for Responsible Healthcare AI

Version

White Paper v14.0 View Log

Date

August 2026

Evidence Status

Conceptual; empirical validation pending

Primary Implementation

Sector Agnostic (Medical Guild Pilot)

Regulatory Status

Non-statutory and non-regulatory unless separately established under applicable law.

Research Note: This document represents an active research initiative. Further versions will be released as the framework evolves through empirical pilots and ongoing structural critique.

Abstract

Artificial intelligence is increasingly embedded within institutions and professional domains in which the consequences of AI outputs cannot be evaluated solely through technical performance. A model may achieve high benchmark accuracy and nevertheless be inappropriate for a particular clinical workflow. A technically reliable system may operate outside an engineering safety envelope. A legally sophisticated system may generate consequences that remain unacceptable to affected communities. An educational system may perform well on generic metrics while undermining the capabilities it is intended to support.

The central governance problem is therefore not simply whether AI systems are technically capable. It is whether the institutions deploying them can determine what the system is authorized to do, and who is entitled to judge its consequences. Governance must define who bears responsibility, how conflicts are managed, and how financial dependencies are disclosed. It must also ensure standards remain contestable, decisions can be reversed, and the governing institution itself remains accountable.

CODE-AI—Coalition of Domain Experts for Artificial Intelligence—proposes a federated institutional architecture for collective AI development, evaluation, governance, and stewardship. Its central Proposal is that where AI creates domain-specific consequences, legitimate domain expertise should have a defined and bounded role in governing those consequences. Its corresponding constitutional constraint is that expertise does not itself create authority, and authority must never become sovereignty.

CODE-AI separates capability, contribution, expertise, authority, ownership, funding, governance, and accountability. The architecture contains Domain Guilds, Technical Commons, and Evaluation Commons. It integrates Public Interest participation and a Constitutional Council to ensure procedural integrity. Supporting these are financial stewardship mechanisms, an operational Secretariat, and a strictly defined Authority Matrix. Finally, it embeds independence safeguards, portability provisions, contestability mechanisms, an Enrollment and Verification Protocol, and a formal research program.

CODE-AI does not propose to replace governments, courts, statutory regulators, professional licensing authorities, institutional governance bodies, or legally accountable decision-makers. Its purpose is narrower: to provide an institutional constitutional layer through which domain expertise, authority, independence, consequence, and accountability can be made explicit in the AI lifecycle.

The architecture is deliberately designed so that its foundational Hypotheses can be materially weakened, unsupported, or falsified under predefined criteria. Its success cannot be established by institutional adoption alone. The relevant question is whether CODE-AI can operate without becoming a cartel, shadow regulator, professional protection mechanism, funding-dependent institution, bureaucratic bottleneck, or self-perpetuating governing elite.

Executive Summary

The rapid deployment of artificial intelligence into societal infrastructure demands a transition from purely technical oversight to robust institutional governance. CODE-AI is a proposed institutional architecture designed to embed specialized domain expertise into the evaluation and governance of AI systems. Recognizing that computational inference is distinct from human professional judgment, CODE-AI systematically bounds authority within defined jurisdictions to prevent technical capability, financial contribution, or professional status from becoming unaccountable sovereign power.

The proposed architecture distributes governance across Domain Guilds, Technical and Evaluation Commons, a Constitutional Council, and a consequence-triggered Public Interest Mechanism. By proposing a strict separation between funding and governance, CODE-AI seeks structural independence. Through mechanisms such as the Funding Concentration Index (FCI) and the Open Standards Covenant, the coalition aims to manage inherent risks of institutional capture, professional cartelization, and antitrust violations.

Evaluated conceptually as a sector-agnostic framework (with specific empirical validation pathways like the Medical Guild pilot — launched at the scale of a single specialty council rather than the full Guild), CODE-AI is presented not as a validated solution, but as a formal, testable Hypothesis regarding how collaborative, domain-participatory governance may safely manage the integration of artificial intelligence into high-consequence human environments.

Key Summary Highlight

CODE-AI is a Hypothesis that structural separation of powers in AI development—dividing capability, funding, expertise, and authority—can create safer, more accountable AI systems without replacing existing legal frameworks.

Clarification of Core Intent

To ensure absolute clarity on what CODE-AI is proposing, the fundamental intention of this coalition must be explicitly stated to correct common misinterpretations regarding "consensus" and "profit."

1. Structured Dissent, Not Forced Consensus

The intention of this coalition is not to force an artificial consensus among all domain experts. In pluralistic societies and complex scientific fields, legitimate disagreement is healthy and necessary. CODE-AI is designed to create a structured, transparent, and accountable environment where experts can rigorously evaluate AI—and where minority scientific positions, risk tolerance differences, and documented dissents are recorded and protected rather than suppressed.

2. A Structural Firewall Against Profit

While AI development, evaluation, computing, and monitoring inherently require vast capital—and commercial entities will inevitably remain major participants—CODE-AI establishes a strict institutional firewall. The claim is not that AI can exist entirely outside of commercial economies. The claim is that profit incentives must never dictate safety-critical governance, evaluation criteria, or the legitimacy of deployment decisions.

Evidence Discipline

CODE-AI distinguishes institutional design propositions from claims about the external world. External factual, empirical, legal and historical claims shall be supported by appropriately authoritative sources. Where evidence is absent, contested or insufficient, the proposition shall not be presented as established fact. The use of a citation shall not imply that the cited source endorses CODE-AI or its institutional architecture.

Claim Status	Language
Established	"Evidence demonstrates..." or "Applicable law establishes..."
Observed	"A documented precedent demonstrates..."
Inference	"The evidence supports the inference that..."
Proposal	"CODE-AI proposes..."
Hypothesis	"CODE-AI hypothesizes..."

This distinction is substantive, not merely editorial. CODE-AI does not claim that its architecture has been empirically demonstrated to improve safety, patient outcomes, institutional legitimacy, independence, or AI performance. Its central mechanisms are Proposals. Its anticipated effects are Hypotheses. Its initial pilots are intended to generate evidence capable of supporting, modifying, or falsifying those Hypotheses.

Introduction

The conventional representation of artificial intelligence development is frequently approximated through a linear technical pipeline: Data → Model → Application → User. This representation is useful for understanding technical development, but it is fundamentally insufficient for understanding consequential deployment.

AI systems increasingly operate inside hospitals, financial institutions, engineering systems, government agencies, educational organizations, scientific research laboratories, legal services, infrastructure, manufacturing, and cybersecurity networks. Within these environments, AI outputs continuously interact with specialized domain knowledge, professional standards, institutional responsibilities, legal duties, resource constraints, human judgment, organizational incentives, and real-world consequences.

The Real-World Socio-Technical Continuum

Click to enlarge

graph LR; A[Domain Need] --> B[Knowledge]; B --> C[Data]; C --> D[Model]; D --> E[Evaluation]; E --> F[Governance]; F --> G[Deployment]; G --> H[Human Decision]; H --> I[Consequence]; I --> J[Feedback]; J -.->|Iterates| A; classDef default fill:#f1f5f9,stroke:#94a3b8,stroke-width:1px,color:#0f172a; classDef governance fill:#e0f2fe,stroke:#38bdf8,stroke-width:2px; class F,H,I governance;

The resulting socio-technical system is therefore more accurately represented as the continuum above. The additional elements in this continuum are not merely workflow steps; they represent complex sources of institutional authority and responsibility. CODE-AI begins from this fundamental observation.

The Governance Problem

AI governance has historically developed through several partially separated traditions. These encompass technical safety, cybersecurity, privacy, fairness, explainability, professional regulation, organizational governance, risk management, and public policy. Each of these traditions addresses important dimensions of the problem, yet a critical unresolved question remains concerning how these dimensions should interact when AI becomes deeply embedded within a consequential domain.

The Crisis of Jurisdiction

The fundamental problem can therefore be expressed as a crisis of jurisdiction: Who governs the consequences of AI when technical capability, domain expertise, institutional authority, and legal responsibility do not reside in the same entity?

This complex governance dilemma is not solved merely by appending an advisory ethics committee to an engineering team. Nor is it solved merely by increasing model accuracy on standardized benchmarks. Furthermore, it is not solved by assuming that the most technically sophisticated organization possesses the inherent legitimacy to determine all downstream societal consequences. CODE-AI proposes that governance must explicitly distinguish different forms of authority and assign them strictly according to defined jurisdictions.

From AI Performance to Institutional Governability

If the metrics of AI governance remain confined to technical performance, the resulting regulatory mechanisms will systematically fail to address institutional reality. A model may demonstrate exceptional statistical accuracy while simultaneously destabilizing a clinical workflow or violating a foundational professional duty. The transition from monitoring performance to supporting governability demands that institutions establish formal mechanisms capable of interpreting computational outputs through the lens of domain-specific responsibilities. Institutional governability requires a system wherein algorithmic actions remain perpetually subject to identifiable human accountability, and wherein the institutions enforcing that accountability remain transparent and structurally independent.

Constitutional Intelligence as the Philosophical Substrate

CODE-AI operates within a broader intellectual proposition: Artificial intelligence is rapidly evolving into a new form of institutional power. Power historically creates a requirement for legitimacy, and institutional legitimacy inevitably requires constraints. Therefore, the emergence of AI-enabled institutional power necessitates constitutional forms of governance.

This is the conceptual territory of Constitutional Intelligence. However, CODE-AI is not itself Constitutional Intelligence. Rather, the relationship functions as a conceptual hierarchy wherein Constitutional Intelligence acts as the philosophical substrate, BeResponsibleAI serves as the broader doctrinal ecosystem, and CODE-AI manifests as the specific institutional architecture for domain-governed AI. This deliberate separation prevents CODE-AI from becoming a universal umbrella containing every other framework, maintaining its distinct operational focus.

CODE-AI: Definition and Scope

The legal and regulatory status of CODE-AI requires extreme precision. CODE-AI is a proposed institutional architecture, a voluntary governance model, a standard-setting and evaluation coordination Proposal, and a live research program. It is potentially implementable through voluntary institutional arrangements.

What CODE-AI Is

Proposed institutional architecture
Voluntary governance model
Standard-setting & coordination Proposal
Live research program

What CODE-AI Is NOT

A government regulator
A statutory authority
A professional licensing body
A substitute for applicable law

Clarification of Terminology: Within CODE-AI, "constitutional authority" refers exclusively to internal contractual or institutional authority governing compliance with CODE-AI's own charter and procedures; it does not constitute constitutional public law.

If CODE-AI mechanisms eventually acquire practical market influence through voluntary adoption, this represents an institutional or market-power consideration. It does not magically transmute voluntary standards into statutory authority. Where legal propositions are made, they remain subordinate to actual applicable law and authoritative legal sources. A CODE-AI body may possess institutionally delegated authority within the CODE-AI architecture without possessing any statutory or sovereign legal authority.

Decision Effects and Limits

To prevent CODE-AI from oscillating ambiguously between a voluntary protocol and a shadow regulator, the exact operational limits of its decisions must be explicitly defined before pilot execution.

Question	What the Document States
What does a Guild produce?	A recommendation, conditional assessment, evaluation report, or other defined output—not a vague "decision."
Who acts on it?	A named deploying institution or accountable decision-maker, not CODE-AI acting as a regulator.
What happens after disagreement?	Dissent record, reconsideration, appeal, escalation, and final accountable institutional decision.
What is public?	Evaluation scope, conflicts, rationale, dissent, conditions, review date, and outcome—subject to legitimate confidentiality limits.
What is enforceable?	Contractual or membership commitments only, unless an external legal authority independently makes it binding.

Constitutional Principles

CODE-AI advances five linked propositions that form its conceptual bedrock:

Domain consequences require domain participation:

Where AI creates specialized consequences, relevant domain expertise should participate in defining evaluation criteria, identifying failure modes, and assessing contextual fitness.

Expertise is not sovereignty:

Domain experts do not automatically acquire legal, institutional, or constitutional authority merely because they possess expertise.

Authority must be explicit:

Every authority-bearing body should have a defined jurisdiction, scope, decision rights, review mechanism, and accountability pathway.

Independence is structural:

An institution cannot establish independence merely by declaring itself independent.

Governance institutions must themselves remain governable:

CODE-AI must be subject to the same principles it proposes for AI governance, including transparency, contestability, bounded authority, attribution, review, correction, portability, and accountability.

Inference is NOT Judgment

A foundational distinction preserved throughout this architecture is the Proposal that inference is not judgment.

It is Observed that regulatory agencies explicitly differentiate between software output and clinical judgment. The U.S. Food and Drug Administration (FDA) issued final guidance in September 2022 on Clinical Decision Support (CDS) Software, which Establishes criteria indicating that software functions intended to enable healthcare professionals to independently review the basis for recommendations—thereby not replacing human clinical judgment—are not regulated as medical devices under specific statutory definitions¹. While this source Establishes a narrower regulatory manifestation of the distinction within a single jurisdiction, CODE-AI treats "inference is not judgment" as a broader normative design proposition. Computational inference should not silently acquire decision authority merely because it is technically capable of producing a statistical answer.

Institutional Architecture

CODE-AI consists of three interacting planes that deliberately prevent the consolidation of authority into a single unified hierarchy: The Commons, The Constitution, and The AI Lifecycle. The core institutional components include:

1. Domain Guilds

Domain-specific bodies responsible for domain expertise and domain-consequence governance.

2. Technical Commons

A shared technical governance and evaluation environment.

3. Evaluation Commons

Shared evaluation infrastructure and methodologies.

4. Constitutional Council

A constitutional and procedural body concerned with institutional integrity, jurisdiction, due process, and constitutional consistency. It is strictly not a super-expert scientific body.

5. Public Interest Mechanism

A consequence-triggered mechanism for representing or investigating credible public-interest concerns.

6. Financial Stewardship

Responsible for transparency, diversification, and structural financial safeguards.

7. Operational Secretariat

Dedicated to administrative and operational execution. Must not be presented as a sovereign governing body.

Federated Governance Flow

Click to enlarge

Domain Guilds

A Domain Guild serves as the principal domain-specific governance body within the CODE-AI architecture. Examples could eventually include a Medical Guild, Engineering Guild, Legal Guild, Financial Guild, Education Guild, Agricultural Guild, or Infrastructure Guild.

The Guild is not simply a passive expert panel. Its institutional purpose is to formalize domain participation in problem definition, domain requirements, evaluation protocols, contextual risk identification, deployment conditions, incident interpretation, and governance learning. However, a Guild's expertise does not automatically create legal authority. Its authority is entirely constructed and defined through the Authority Matrix and through whatever institutional or contractual arrangements legitimately recognize its subsequent decisions.

Technical Commons

The Technical Commons provides the engineering expertise, interoperability standards, reproducibility mechanisms, infrastructure knowledge, security protocols, testing tools, and technical implementation support necessary for robust AI evaluation. Crucially, the Technical Commons does not acquire domain authority simply because it provides the underlying technical substrate.

It is Observed that successful open technical commons models rely on strict procedural governance to separate technical contribution from unilateral control. For example, the World Wide Web Consortium (W3C) explicitly utilizes structured consensus, public review stages, implementation experience requirements, and membership approval within its standard-setting processes, paired with a patent policy designed to support the implementation of specifications on a royalty-free basis². This precedent illuminates a specific governance mechanism for managing technical standards, though it does not empirically validate the entire CODE-AI architecture.

Evaluation Commons

The Evaluation Commons provides the shared infrastructure necessary for benchmarks, testing protocols, adversarial evaluation, reproducibility, domain-specific test suites, incident analysis, and comparative evaluation. It is explicitly not intended to finance frontier-model development. Its purpose is to reduce the duplication of effort and to preserve a vital structural independence between the model developer and the model evaluator. Where absolute independence is practically impossible, the precise nature of the relationship must be transparently disclosed and subjected to rigorous governance controls.

Authority Matrix

The Authority Matrix proposes explicitly assigning the type and boundaries of authority, so that decision rights are categorized rather than treated as a single, homogeneous quantity.

Authority Type	Principal Question
Domain authority	What constitutes acceptable domain performance or consequence?
Technical authority	What is technically feasible, valid, or reproducible?
Procedural authority	Was the governance process followed correctly?
Legal authority	What does applicable law require or permit?
Organizational authority	Who has the institutional power to authorize deployment?
Public-interest standing	Who may trigger review or bring affected interests into the process?
Constitutional authority	Whether CODE-AI's own rules were followed?

Crucial Safeguard: These authorities must not be conflated. A Constitutional Council should not assume super-expert scientific status and overrule a medical determination merely because it possesses constitutional oversight authority. Likewise, a Technical Commons should not unilaterally dictate patient acceptability criteria merely because it understands the underlying model architecture. A Domain Guild must never determine statutory legality merely because it comprehends the nuances of a professional context.

Constitutional Council

The Constitutional Council functions as the procedural and constitutional referee of CODE-AI. It is tasked with determining whether authority was exercised within assigned jurisdictions, whether conflicts were disclosed, whether required processes were followed, whether standing was improperly denied, or whether a constitutional rule was violated. It should generally not decide substantive scientific questions, such as which medical hypothesis is factually correct. The Council is a guardian of process, not a universal adjudicator of empirical truth.

To prevent the Constitutional Council from devolving into an unaccountable ruling elite, CODE-AI proposes separating initial formation from transitional governance, utilizing distributed multi-constituency appointments, enforcing strict term limits, implementing defined removal safeguards, and requiring conflict disclosures. Its jurisdiction is strictly published and bounded. Reviews are bound by a "reasonableness" standard of procedural adherence and must be completed within defined time limits. The Council must publish reasoned opinions for all rulings, and a sunset clause requires all delegated powers to automatically expire every 24 months unless renewed by evidence-based review.

Where a challenge concerns the Council's own conduct rather than a Guild's or Secretariat's, review is conducted by members drawn from constituencies not party to the dispute, so that the body accused of a constitutional breach is never the sole judge of that breach.

Public Interest Mechanism

Public-interest participation is necessary where consequence thresholds are met, particularly when AI consequences extend far beyond the walls of participating institutions. However, unrestricted public-interest authority risks devolving into procedural weaponization. CODE-AI therefore carefully distinguishes between participation, standing, complaint, investigation, review, decision, and veto.

The Proposal dictates a consequence-triggered model. The mechanism becomes active when credible evidence indicates potentially material consequences to affected populations. A trigger requires a defined threshold, evidence or credible grounds, conflict disclosure, and procedural review. This design explicitly includes safeguards against frivolous complaints, bad-faith challenges, competitor-sponsored procedural obstruction, and repeated delays. It is an investigative and representative body; it is not a permanent public-interest veto.

To avoid "echo chamber" isolation, CODE-AI proposes that ethicists, sociologists, behavioral scientists, and patient advocates are allocated full voting seats within every Guild-level pilot from its outset, rather than being granted merely observer status or deferred to a later phase of the coalition's growth. To prevent "token representation," these seats must be defined by fixed quotas, subject to specified minimum participation requirements, explicit conflict rules, structured compensation frameworks, and the guaranteed right to publish formal dissents to any technical evaluation. Participants within this mechanism are explicitly shielded by a strict anti-retaliation policy.

Financial Stewardship

Financial dependence can inexorably produce influence, even in the total absence of formal voting rights. The architecture therefore proposes a strict separation between financial contribution and governance authority. A funder may provide capital, compute, equipment, datasets, infrastructure, or technical expertise. None of these contributions should automatically purchase appointment rights, voting rights, domain authority, constitutional authority, or control over evaluation outcomes.

It is Observed that the Apache Software Foundation (ASF) provides an institutional precedent for decoupling funding from technical governance. The ASF explicitly notes that sponsors are not part of ASF corporate governance and that becoming a sponsor does not confer any organization or its employees specific merit within the ASF or its projects³. While this structure does not intrinsically prove that the identical mechanism solves complex clinical safety governance, it Establishes the institutional feasibility of separating sponsorship from decision-making authority.

The Financial Stewardship Function is responsible for maintaining transparency, monitoring diversification, and enacting structural financial safeguards. At formation, CODE-AI will publish a baseline dependency register documenting monetary funding and material in-kind support, including volunteer labour, institutional affiliation, hosting, software, compute, data access, legal support, administrative support, and any founder-provided resources. A zero monetary funding score does not itself establish operational independence.

Operational Secretariat

The Operational Secretariat exists to execute administrative duties. It provides administration, documentation, coordination, financial reporting, record-keeping, technical infrastructure, meeting support, and implementation assistance.

A central CODE-AI Proposal is that information custody must not become decision sovereignty. The Secretariat must not acquire substantive authority merely because it controls the flow of information. It must therefore be subjected to rigorous audit, disclosure requirements, access controls, separation of duties, and continuous constitutional oversight.

Independence Architecture

CODE-AI fundamentally rejects the assumption that institutional independence can be achieved through mere declaration. It proposes a Four-Dimension Independence Test, requiring rigorous assessment across:

1. Appointment Independence

Who selects the people who hold authority?

2. Decision Independence

Can decision-makers act without improper direction?

3. Financial Independence

Can the institution remain operational without becoming dependent upon a dominant funder?

4. Removal Independence

Can an interested party remove a decision-maker because of an unfavorable decision?

These dimensions are entirely distinct. An organization might achieve appointment independence while completely lacking financial independence. It is therefore a multidimensional Hypothesis that measuring and enforcing these distinct vectors improves the overall legitimacy of the institution. CODE-AI further proposes that this same four-dimension test applies recursively within each Domain Guild, and not only at the level of the coalition as a whole — a Guild that is internally appointment-independent, decision-independent, financially independent, and removal-independent of the coalition's funders may nonetheless fail all four dimensions with respect to its own members' individual financial relationships, a distinct and equally material vulnerability addressed in the following section.

Conflict of Interest

CODE-AI does not operate under the naive assumption that expert institutions are conflict-free. In the realm of professional and scientific governance, financial conflicts of interest are deeply entrenched.

It is Established by a systematic review of 37 studies encompassing 14,764 clinical guideline authors that 45% possessed at least one financial conflict of interest (FCOI)⁴. The prevalence of these conflicts varies significantly by specialty and sponsoring society, with profound implications for the objectivity of standard-setting. Further empirical scrutiny underscores the inadequacy of voluntary disclosure. In a cross-sectional analysis of 20 clinical practice guidelines comparing author self-disclosures against the Open Payments database, it was Observed that while only 37.4% of authors self-disclosed industry relationships, a staggering 73.7% were actually found to have received industry payments⁵. Furthermore, 81.8% of the panel chairs evaluating these guidelines possessed financial conflicts, none of whom disclosed them accurately. The same analysis found inaccurate disclosure to be significantly more common among professors than non-professors — a finding directly relevant to Guild recruitment, since founding cohorts naturally gravitate toward senior figures for institutional legitimacy, and that is evidently not the population carrying the lowest conflict risk. Similar conflict prevalence has been documented across various specialties, including hematology clinical practice guidelines⁶.

The appropriate Inference is not that medical experts are inherently compromised or perpetually biased; rather, the evidence dictates a much narrower conclusion: Financial conflicts exist at highly meaningful rates in standard-setting, and therefore, institutional mechanisms must be specifically engineered to identify, disclose, and manage them.

CODE-AI proposes that individual conflict management does not equal structural institutional independence. Recusal is a necessary but entirely insufficient mechanism. A person may dutifully recuse themselves while the broader institution remains structurally dependent upon a particular funder or professional constituency.

Professional Self-Governance and Capture Risk

Professional self-governance serves as a vital institutional precedent, yet it represents a profound area of vulnerability requiring stringent structural safeguards. Domain expertise is strictly necessary for governing domain consequences, but expertise does not automatically confer unlimited governance authority.

Professional capture must remain an explicit design consideration. A Domain Guild could theoretically reject an advanced AI system not because it poses a genuine risk to public safety, but because it threatens professional income, disrupts existing lucrative workflows, diminishes institutional prestige, or undermines incumbent business models. CODE-AI therefore rejects the premise that "human authority" acts as a legitimate proxy for the "protection of existing professional interests." The Guild must systematically distinguish between verifiable patient or public safety concerns and the mere preservation of the profession. This distinction must be empirically evaluated rather than assumed.

Funding Concentration

CODE-AI seeks to meticulously preserve the distinction between financial diversification and true institutional independence. Diversification alone does not automatically guarantee independence, and CODE-AI does not invent arbitrary funding thresholds unless explicitly supported by institutional necessity.

To make dependency mathematically visible, CODE-AI proposes a Funding Concentration Index (FCI). The FCI operates strictly as an internal governance metric, not as an established regulatory standard.

Funding Concentration Index (FCI)

$$ FCI = \sum_{i} (s_i \times 100)^2 + P_c + P_r $$

$ s_i $ = The decimal share of total funding from source i (e.g., 40% = 0.40, making $ 0.40 \times 100 = 40 $, yielding $ 40^2 = 1600 $).
$ P_c $ = Compute dependency penalty (e.g., +500 applied if infrastructure relies on a single dominant commercial vendor).
$ P_r $ = Restricted-funds penalty (e.g., +500 applied for earmarked corporate funding).

Recalibrated Thresholds & Worked Example

Example: A diversified funding split of 40%, 30%, 20%, and 10% (with no penalties) yields an FCI of $ 1600 + 900 + 400 + 100 = 3000 $, placing it safely in the 'Moderate / Scrutiny' tier.

FCI Score	Independence Status
< 2500	Highly Diversified
2500 – 4000	Moderate / Scrutiny Required
> 4000	Concentrated / Failed Independence

CODE-AI proposes a declining ceiling on any single funding source's share of annual operating budget during the founding years. This moves toward a materially lower steady-state ceiling as the funding base diversifies, published and audited alongside the FCI itself.

When examining nonprofit structures, CODE-AI carefully distinguishes between legal requirements and its own governance Proposals. For instance, U.S. tax law Establishes a "public support test" (under IRC Sections 509(a)(1) and 170(b)(1)(A)(vi)) which calculates ratios of public contributions over a five-year period to determine nonprofit classification⁷. This represents a specific jurisdictional tax-law classification, not a universally ideal constitutional limit on funder concentration. Legal public-support thresholds and CODE-AI's internal institutional independence metrics remain conceptually distinct.

Data, IP and Portability

The original, romanticized concept of open-source "forking" or "exit" is fundamentally inadequate in the context of capital-intensive AI governance if the departing institution lacks the practical means to operate independently.

CODE-AI Proposes Institutional Portability: Portability is the practical constitutional safeguard; institutional exit is a consequence of portability where lawful and feasible. A governance system should actively avoid making meaningful participation dependent on an irreversible loss of institutional alternatives, wherever technically and legally feasible.

Portability may involve benchmark specifications, evaluation protocols, schemas, interfaces, documentation, model cards, provenance metadata, reproducible procedures, synthetic test datasets, licensing rights, and legally transferable datasets. However, because an organization cannot necessarily take protected clinical data, proprietary model weights, licensed datasets, confidential information, or restricted compute resources with it, portability explicitly does not guarantee unrestricted exit. No constitutional "right to exit" automatically allows copying or transferring legally protected assets.

Open Standards and Antitrust Safeguards

When competitors and professional standard-setting organizations collaborate, they generate antitrust risks. Under U.S. federal antitrust law, it is Established that professional associations are not automatically shielded from antitrust scrutiny, and the Federal Trade Commission (FTC) frequently takes action against association rules that illicitly restrict competition.

The U.S. Supreme Court Established in American Society of Mechanical Engineers, Inc. v. Hydrolevel Corp. (456 U.S. 556 (1982)) that a nonprofit standard-setting association can be held civilly liable for treble damages arising from Sherman Antitrust Act violations — the treble-damages civil remedy itself being provided under Section 4 of the Clayton Act — when its volunteer agents, acting with "apparent authority," manipulate safety codes to disadvantage a competitor⁸. In that case, committee members affiliated with a dominant corporation drafted an "unofficial" ASME response declaring a competitor's product unsafe, effectively weaponizing the association's reputation to destroy a rival. The Court concluded that a rule imposing liability on standard-setting organizations is necessary because they wield great economic power and are best situated to prevent antitrust violations achieved through the abuse of their formidable reputations.

Furthermore, in North Carolina State Board of Dental Examiners v. FTC (574 U.S. 494, 135 S. Ct. 1101 (2015)), the Supreme Court Established that a state regulatory board controlled by active market participants requires "active supervision" by the State to successfully invoke state-action antitrust immunity⁹. The Inference is clear: when industry experts govern their own domains, the risk of cartelization and exclusionary behavior is acute, necessitating rigorous, independent procedural supervision.

CODE-AI is legally and institutionally aware that standards and collective governance structures can acquire exclusionary effects. Therefore, it proposes the adoption of an Open Standards Covenant incorporating safeguards such as openness, non-discriminatory participation, transparent criteria, procedural contestability, explicit appeal mechanisms, and the avoidance of unnecessary barriers to entry. The separation of genuine safety governance from commercial self-interest is paramount. These are presented as CODE-AI design proposals informed by antitrust principles, not as universal legal requirements unless directly supported by applicable law.

Governance Latency and Proportionality

AI development operates at a uniquely high velocity. If governance processes are misaligned with this speed, they become practically irrelevant. CODE-AI explicitly treats governance latency as a measurable design problem rather than an inevitable bureaucratic reality. A governance process must rigorously record time to review, time to decision, time to appeal, time to escalation, time to remediation, and time to constitutional review.

It is Observed that successful standards organizations utilize specialized methodologies to balance deliberation with necessary forward momentum. The Internet Engineering Task Force (IETF) relies on the principle of "rough consensus and humming" (as detailed in RFC 7282) to avoid the paralyzing delays of absolute unanimity and the polarizing effects of formal voting¹⁰. The W3C similarly pairs consensus mechanisms with explicit procedural stages, public review, and implementation experience.

Proportionality Equation

$$ \text{Governance Intensity} \propto (\text{Consequence} \times \text{Risk} \times \text{Uncertainty} \times \text{Velocity}) $$

Each factor is scored on a predefined 1–5 scale by the Operational Secretariat during initial intake. Any participant may trigger a challenge to this score, which is resolved by the Constitutional Council.

Score Product	Required Governance Response
1 – 25	Ordinary Review
26 – 100	Enhanced Documentation Required
101 – 400	Mandatory External Audit
401 – 625	Suspension or Emergency Review

CODE-AI proposes that governance must be proportionate to consequence and sufficiently responsive to the pace of the systems it governs. This does not imply that high-risk systems receive less scrutiny; it proposes that scrutiny should be structurally designed so that high-risk governance remains operationally usable.

Legal and Regulatory Boundaries

The boundaries of accountability and liability within CODE-AI must not create legal ambiguity. Participation in the CODE-AI ecosystem does not automatically transfer professional liability, product liability, institutional liability, regulatory liability, or statutory responsibility away from the legally accountable deploying entities.

Where a CODE-AI assessment contributes to an organizational deployment decision, the applicable law determines the ultimate legal consequences. The architecture strictly preserves the distinction between a governance recommendation and legal responsibility. As established by the Final Responsibility Clause: Every pilot activity shall identify a named legally accountable institution or person responsible for the resulting action. No CODE-AI output shall be interpreted as transferring, limiting, or extinguishing that responsibility. This critical distinction must be reflected in all contracts, policies, and deployment documentation. Furthermore, CODE-AI does not imply that it can override national law, statutory regulators, courts, professional licensing laws, or binding institutional legal obligations.

Medical Guild: A Defensible Proof of Concept

Because foundation models (like those from OpenAI, Anthropic, or Google) are general-purpose, their outputs impact virtually every sector—law, finance, education, infrastructure, and coding. Therefore, the ultimate CODE-AI governance architecture must inevitably be cross-domain and sector-agnostic.

However, healthcare is a defensible and strategically useful starting point because the stakes (human life and safety) are universally understood, and the tension between statistical inference and clinical judgment is already a recognized regulatory issue.

If the Medical Guild pilot demonstrates feasibility and acceptable trade-offs, CODE-AI may test whether selected governance primitives—such as bounded authority, conflict controls, transparent dissent, and evaluative independence—can be adapted to other domains. Cross-domain transfer must be empirically assessed, not presumed.

Scaling Limitation: CODE-AI proposes that the Medical Guild not be launched as a single, fully populated body. Its first operating unit should instead be a single specialty council — for example, a Radiology AI Council — convened at a scale (approximately fifteen to thirty participants) small enough for direct, personal vetting of each founding member's practice status, institutional affiliation, and disclosed conflicts, without first requiring the coalition to build external credential-verification infrastructure. Expansion to the full Medical Guild, and subsequently to additional Domain Guilds, is a decision to be made from the pilot's evidence, not a precondition assumed in advance.

Founding Cohort Disclaimer: The founding Radiology AI Council is a time-limited feasibility-research cohort, not a representative professional body, regulator, accreditor, certification panel, procurement committee, vendor-ranking body, or market-facing standards authority. During the pilot, its outputs are limited to documented research observations, governance analyses, and non-binding advisory assessments. They do not constitute clinical advice, product clearance, legal advice, procurement recommendations, endorsements, or determinations of market access.

Research Program

The CODE-AI implementation strategy functions concurrently as an institutional research program.

The Central Hypothesis

Structurally independent, domain-participatory governance may improve the governability and safety-relevant properties of AI systems deployed in consequential domains compared with governance arrangements characterized by greater dependency or weaker domain participation.

The initial pilot will investigate fundamental feasibility, operability, independence, contestability, transparency, conflict management, portability, and governance latency. A single pilot cannot establish universal superiority, causal improvement in patient outcomes, optimal funding concentration, or the elimination of professional and corporate capture. After achieving feasibility, the research program will advance to comparative studies examining Model A (CODE-AI-governed deployment) against Model B (Conventional governance).

Because CODE-AI is evaluating itself, the research program must maintain strict independent research governance. This requires external Institutional Review Board (IRB) or equivalent ethics approval before commencement. Furthermore, raw evaluation data must be held under independent stewardship, and all material adverse findings, failed hypotheses, or protocol deviations must be published (see Appendix A: Negative-results clause).

Measurement Framework

CODE-AI explicitly avoids reliance on a single, reductive "CODE-AI score." Instead, the objective is to create an empirical, auditable record.

Metric Category	Principal Question
Authority concentration	How concentrated are decision rights?
Funding concentration	How dependent is the institution on major funders?
Appointment concentration	How concentrated is selection power?
Conflict burden	How many decisions involve declared conflicts?
Contestability	How frequently can decisions be challenged?
Resolution quality	How often do challenges result in correction?
Governance latency	How long does decision-making take?
Portability	Can critical assets actually be reproduced or transferred?
Bypass rate	How often do participants circumvent governance mechanisms?
Capture indicators	Are decisions systematically aligned with a particular constituency?
Outcome indicators	Do governed systems demonstrate better safety-relevant outcomes?

Structural Self-Critique and Vulnerabilities

A robust governance architecture must actively map its own blind spots. Rather than awaiting external validation, CODE-AI internalizes several critical structural critiques to stress-test its own resilience and identify necessary evolutionary refinements.

The "Echo Chamber" Vulnerability

Focus: Sociotechnical & Interdisciplinary Blindspots

The Internal Critique: The proposed "Domain Guilds" risk devolving into narrow echo chambers comprised solely of technologists and specialized clinicians, inadvertently ignoring the broader downstream societal and human impacts of AI deployment.
Architectural Refinement: CODE-AI must explicitly mandate interdisciplinary representation within the Guilds. True human-centered governance requires formally allocating full voting seats—not merely observer status—to ethicists, sociologists, behavioral scientists, and patient advocates to accurately map downstream consequences.

The Compute Dependency Threat

Focus: Infrastructure & True Independence

The Internal Critique: Compute dependency is a major and measurable threat to evaluative independence, requiring disclosure, diversification where feasible, contractual safeguards, reproducibility pathways, and contingency planning.
Architectural Refinement: Full hardware independence may be desirable in some contexts, but it should not be presented as a prerequisite for any credible pilot. Instead, the focus must be on mitigating hardware lock-in through transparency rather than assuming we can bypass the commercial compute market entirely.

Falsifiability

A serious institutional hypothesis must explicitly specify its failure conditions. Falsifiability is not a hostile critique; it is a foundational component of the research methodology. CODE-AI would be considered materially weakened, unsupported, or falsified under predefined criteria if empirical implementation demonstrated that:

Condition for Falsification	Implication
Failed Independence	Structural safeguards and funding diversification fail to measurably reduce dependency.
Cartelization	Guilds demonstrate systematic, exclusionary professional protectionism.
Illusion of Exit	Portability proves economically or technically meaningless in practice.
Procedural Weaponization	Public-interest mechanisms are systematically weaponized by bad-faith actors or competitors.
Paralysis	Constitutional procedures create unacceptable governance latency preventing timely deployment.
Market Exclusion	CODE-AI devolves into an exclusionary market-access monopoly.
Information Capture	The Secretariat accumulates substantive authority through the unchecked control of information.
Unaccountable Elite	The Constitutional Council becomes an unaccountable political body.
No Measurable Benefit	Implementation produces no measurable benefit compared with conventional, unilateral corporate governance.
Routine Bypass	Participants routinely bypass the architecture because its bureaucratic costs exceed its institutional benefits.

These are not rhetorical concessions; they are rigorous empirical failure conditions.

Limitations

CODE-AI possesses substantial, unresolved limitations that form the basis of its long-term research agenda. The architecture introduces institutional complexity that may be significantly more burdensome than conventional, top-down governance. Meaningful domain participation requires extensive time and highly specialized expertise, which may be difficult to sustain. Funding diversification mitigates, but cannot entirely eliminate, deep-seated resource dependency. Both expert capture (where professionals protect their own interests) and corporate capture (where industry leverages funding for influence) remain persistent, existential threats.

Furthermore, constitutional procedures inevitably introduce governance latency, which may conflict with the rapid deployment cycles of frontier AI. The legal consequences of voluntary institutional standards remain highly uncertain and highly dependent on specific jurisdictional implementations. Finally, true portability is severely constrained by data protection laws, intellectual property rights, localized compute requirements, and complex contractual obligations, making the promise of "exit" inherently incomplete.

Implementation Roadmap

CODE-AI proposes a staged Evidence Maturity Model, ensuring that adoption alone is not mistaken for scientific validation. The staged model — 100 founding participants, then 1,000, then 10,000, then 100,000, then, at network maturity, approximately 1,000,000 contributors — is a long-run description of the coalition at scale, not a near-term recruitment target. Attempting registration at scale before the constitutional mechanisms below have been load-tested is precisely the sequence the Falsifiability table's Cartelization and Unaccountable Elite conditions warn against; each phase therefore gates the next on evidence, not on a fundraising or membership calendar.

Phase I — Constitutional Formation

Establish the legal host structure, constitutional document, Authority Matrix, appointment rules, financial stewardship rules, conflict-of-interest policy, public-interest mechanism, amendment process, and transparency policy — before any individual is recruited and before any funding is accepted. Concretely, this phase requires:

Incorporation of an actual legal host entity, so that enrollment, disclosure, and funding rules have a body to attach to.
Ratification of the funding glide-path, the Four-Dimension Independence Test, the Enrollment and Verification Protocol (Appendix D), and the Public Interest Mechanism as founding documents, not as amendments added once participants have already joined.
At formation, CODE-AI will publish a baseline dependency register documenting monetary funding and material in-kind support, including volunteer labour, institutional affiliation, hosting, software, compute, data access, legal support, administrative support, and any founder-provided resources. A zero monetary funding score does not itself establish operational independence.

Phase II — Medical Guild Formation

Establish the bounded mandate, membership criteria, domain authority, conflict management processes, dissent procedures, and initial evaluation methodology — beginning with a single specialty council rather than the full Guild. Concretely, this phase requires:

Narrowing the first operating unit to a Radiology AI Council (or comparable single-specialty council), convened within the convening authority's own professional network, at a scale small enough for direct personal vetting.
Recruiting the founding cohort by direct invitation and peer nomination rather than open public registration, since an open call self-selects for availability rather than practising credibility.
Verifying credentials and requiring financial-disclosure declarations as a condition of admission, per Appendix D, rather than deferring disclosure to the point of a specific decision.
Produce concrete operational artifacts (e.g., formal charters, decision dossiers, conflict-management algorithms, and fillable templates to prevent theoretical drift).

III

Phase III — Evaluation Infrastructure

Develop the benchmark registry, evaluation protocols, evidence registry, audit mechanisms, and reproducibility procedures as deployable, actionable instruments.

Phase IV — Pilot (Proof of Concept)

Rigorously test feasibility, governance latency, independence, contestability, portability, conflict management, and decision quality in a controlled environment, over a period of twelve to twenty-four months. Mandatory Prerequisite: Convening this phase requires jurisdiction-specific legal and ethics (Institutional Review Board or equivalent) review.

Phase V — Comparative Research

Compare CODE-AI governance with conventional arrangements where ethically and practically feasible.

Phase VI — Replication

Replicate the architecture in additional professional domains, and only then consider expansion toward Stage 2 of the contributor model, if and only if the empirical evidence from the pilot justifies expansion.

Discussion

CODE-AI should not be evaluated primarily by whether its theoretical institutional architecture appears elegant on paper. The relevant and ultimate test is whether the architecture produces a measurable improvement in the ability of human institutions to govern highly consequential AI.

This requires confronting several deeply uncomfortable possibilities. A technically sophisticated Domain Guild may inevitably devolve into a professional cartel. A meticulously diversified funding model may remain economically dependent on an oligopoly of compute providers. A well-intentioned public-interest mechanism may become a procedural weapon utilized by corporate rivals. An open standard may mutate into a fortified market-access barrier. A Constitutional Council intended as a referee may slowly morph into an unaccountable political elite.

These inherent risks are not reasons to abandon the architecture; they are the exact, compelling reasons to subject the architecture to relentless empirical testing.

Conclusion

Artificial intelligence is rapidly transitioning from being merely a discrete computational tool toward becoming foundational infrastructure embedded within vital human institutions. As AI systems become more consequential, technical performance metrics alone become an alarmingly incomplete basis for governance. The missing, central institutional question is: Who possesses legitimate authority over the real-world consequences of AI?

CODE-AI proposes one potential answer. It proposes a federated institutional architecture in which domain expertise is structurally represented; technical expertise remains distinct; evaluation is systematically separated from sponsorship where feasible; authority is explicitly assigned and bounded; expertise is never treated as equivalent to absolute sovereignty; funding is separated from governance authority; independence is assessed structurally rather than individually; conflicts of interest are managed rather than denied; the public interest is protected through consequence-triggered standing; standards remain procedurally contestable; portability is maximized where lawful; governance intensity is proportional to consequence; and the governing institution remains perpetually subject to challenge.

The architecture is deliberately modest in its claim. It does not claim to have solved the intractable problem of AI governance. It does not claim that professional experts are inherently independent. It does not claim that diversified funding guarantees uncompromised independence. It does not claim that voluntary standards are immune from antitrust risk. It does not claim that portability guarantees a seamless institutional exit. It does not claim that a single sector Pilot will instantaneously improve outcomes. Instead, it proposes a rigorous set of mechanisms through which these vital propositions can be empirically tested.

The most important commitment of CODE-AI is therefore methodological: The architecture must remain relentlessly accountable to evidence regarding its own consequences. Its success should not be defined by its institutional survival or market adoption. Its success must be defined by whether it can empirically demonstrate that bounded, domain-participatory, independent, and contestable governance materially improves the governability of consequential AI without inadvertently creating a new, unaccountable concentration of institutional power.

The ultimate test is therefore recursive. AI must remain governable. The institutions governing AI must remain governable. And the constitutional architecture governing those institutions must remain fully capable of being challenged, amended, or abandoned.

"Intelligence creates capability. Capability creates power. Power requires legitimacy. Legitimacy requires governability. And those who govern power must themselves remain governable."

Appendices

Appendix A — Core Constitutional Rules

Contribution shall not purchase authority.
No funder shall possess unilateral domain or constitutional authority.
Expertise shall not constitute sovereignty.
Conflicts shall be disclosed and managed structurally.
Recusal shall not substitute for institutional independence.
Authority shall be explicitly assigned and bounded.
No institution shall be assumed permanently immune from capture — including a Domain Guild with respect to its own members.
Public participation shall be distinguished from decision authority.
CODE-AI shall not claim statutory authority unless separately established by law.
Standards shall be developed with competition and antitrust safeguards.
Portability shall be maximized where legally and technically feasible.
Governance intensity shall be proportional to consequence, risk, uncertainty, and change velocity.
No individual shall be enrolled as a domain expert without verified practice status and a disclosed-conflicts declaration.

A.1 — Core Pilot Governance Clauses

Pilot Suspension Clause: The pilot shall be suspended when predefined safety, independence, conflict, latency, data-protection, or procedural-integrity thresholds are exceeded. Suspension shall not require a finding of individual fault and shall remain available as a precautionary governance measure.
No Self-Certification Clause: No participating person or institution may be the sole authority for verifying its own credentials, conflicts, independence, performance, or compliance where an independent verification pathway is reasonably available.
Negative-Results Clause: CODE-AI shall publish material adverse findings, failed hypotheses, unresolved disputes, and protocol deviations, subject only to lawful confidentiality, privacy, security, and intellectual-property limitations.
Anti-Retaliation Clause: No participant, evaluator, public-interest representative, or affected person shall suffer exclusion, loss of role, reputational retaliation, or funding retaliation for submitting a good-faith dissent, complaint, conflict disclosure, or safety concern.
Emergency-Action Clause: Emergency action may temporarily shorten ordinary procedures only where delay presents a materially greater risk, and every emergency action shall receive retrospective review within a defined period.
Change-Control Clause: No material change to the pilot's authority, membership composition, evaluation method, funding structure, data practices, or publication obligations shall take effect without versioned documentation, public notice, conflict review, and independent approval.
Public-Interest Access Clause: Affected persons and their authorized representatives shall have an accessible procedure for submitting concerns, receiving acknowledgment, requesting review, and obtaining a reasoned response, subject to lawful confidentiality and security restrictions.
Sunset Clause: Each pilot body, delegation of authority, and exceptional procedural power shall expire automatically unless renewed through a documented review based on evidence.
Founder-Power Clause: No founder, author, host institution, funder, or convening authority shall possess permanent appointment, veto, interpretive, amendment, or information-control rights solely by virtue of founding or supporting CODE-AI.
Data-Governance Clause: The pilot shall specify data ownership, access rights, retention periods, deletion procedures, provenance requirements, privacy safeguards, permitted secondary use, and responsibility for breach notification before data collection begins.

Appendix B — Relationship to BeResponsibleAI and IRHAI

The hierarchical relationship must remain conceptually clear and non-duplicative, representing distinct operational layers of governance:

Constitutional Intelligence: The philosophical and doctrinal substrate regarding intelligence, power, and legitimacy.
BeResponsibleAI: The broader conceptual movement mapping institutional governability.
CODE-AI: The proposed collective institutional architecture for domain-governed AI.
IRHAI (Institute for Responsible Healthcare AI): The specific healthcare-focused institutional and governance environment.
Domain Methodologies: Highly specific functional instruments, including:
- PRIME: Pre-development admissibility assessment.
- RATSe / RATSe-Health: Governability assessment framework.
- PCCM: Consequence and authority mapping architecture.
- DSEA: Deterministic implementation architecture (where appropriate).

CODE-AI does not absorb or redefine these frameworks. It provides the secure institutional environment in which they may be applied, tested, and refined.

Appendix C — Final Epistemic Position

CODE-AI is a Proposal. Its institutional mechanisms are design propositions. Its anticipated benefits are Hypotheses. Its precedents are Observations. Its externally supported factual claims should remain strictly evidence-backed. And its ultimate legitimacy must derive not from the theoretical elegance of its constitution, but from what empirically transpires when that constitution is placed under intense institutional stress. CODE-AI should not ask the world to believe that it works. It should construct the rigorous conditions under which the world can definitively find out.

Appendix D — Enrollment and Verification Protocol

The preceding architecture defines what Domain Guilds do once populated. It does not, by itself, define who is let in or how a claimed expertise is checked before authority is extended. This appendix closes that gap. It applies to every enrollment into a Domain Guild or Council, beginning with the Radiology AI Council pilot described under the Medical Guild Pilot and Phase II of the Implementation Roadmap.

D.1 — Admission Criteria

Verified, active practice status: A credential that once existed is not sufficient; enrollment requires confirmation of current, active practice in the relevant specialty.
Institutional diversity: No single hospital, department, or funder relationship shall hold more than a bounded share of any founding cohort's composition.
Mandatory, cross-checked financial disclosure: Disclosure of financial relationships is required at the point of admission, not only at the point of a specific decision, and is cross-checked against a public record where one exists rather than accepted on a self-report basis alone.
Recruitment by direct invitation and peer nomination: Rather than open public registration, for the founding cohort of each new Guild or Council. Each admitted member may nominate one further candidate, on the model of a fellowship or academic-committee appointment process.

D.2 — Verification Mechanics

Where a government-maintained, publicly searchable professional register exists, CODE-AI proposes using it as the baseline license-verification instrument rather than constructing new verification infrastructure. It is Observed that India's National Medical Commission maintains the Indian/National Medical Register, a government-run, publicly searchable database of over 1.4 million registered medical practitioners by registration number, name, and State Medical Council¹¹. For the Radiology AI Council pilot, CODE-AI proposes that each applicant's National Medical Register entry be checked against their declared registration number as a condition precedent to admission. Equivalent public or professional-body registries should be used as they become the relevant baseline in other jurisdictions or other Domain Guilds.

D.3 — Disqualification and Review

An applicant who cannot be matched to an active entry in the relevant professional register shall not be admitted.
An applicant whose disclosed financial relationships are found, on cross-check, to have been materially incomplete shall be referred for review rather than automatically admitted or automatically excluded, consistent with the Constitutional Council's procedural (not substantive) jurisdiction described above.
Admission criteria and disqualification determinations are themselves subject to the Contestability and Appeal mechanisms described elsewhere in this architecture, and are not delegated solely to the Secretariat.

Appendix E — Standalone Pilot Protocol

This document governs the Phase IV Radiology AI Council Feasibility Pilot. It is distinctly separate from the broader conceptual white paper and establishes the enforceable operating parameters for the founding cohort.

Radiology AI Council: Feasibility Pilot Protocol

Phase IV Operating Document

1. Scope and Cohort Boundary

The founding Radiology AI Council is a time-limited feasibility-research cohort, not a representative professional body, regulator, accreditor, certification panel, procurement committee, vendor-ranking body, or market-facing standards authority. During the pilot, its outputs are limited to documented research observations, governance analyses, and non-binding advisory assessments. They do not constitute clinical advice, product clearance, legal advice, procurement recommendations, endorsements, or determinations of market access.

2. Pilot Conduct & Competition Protocol

Before the first meeting is convened, all participants agree to the following enforceable rules to preempt antitrust and exclusionary risks:

A written statement acknowledging the Council does not certify, approve, endorse, rank, procure, or exclude AI products or vendors.
All agendas must be pre-circulated, and contemporaneous minutes formally recorded.
Prohibited topics: Explicit prohibition on discussions regarding prices, fees, discounts, market allocation, coordinated purchasing, refusal to deal, commercial strategies, or confidential competitor information.
Conflict disclosure is required before every agenda item, accompanied by strict recusal rules and a recorded rationale for every recusal decision.
Fair Process: Any named-product analysis requires a scope notice, a defined evidence standard, an opportunity for the developer to correct factual errors, and a formal response record prior to publication.
Legal review is required prior to the publication of any output that could reasonably be read as adverse to a named developer, product, hospital, or professional group.
All outputs must carry standard wording that defines limitations and prevents reliance as a regulatory or clinical decision.

3. Concrete Success Thresholds

The pilot will be deemed successful and eligible for Phase VI expansion only if all the following empirical thresholds are met:

Review Latency: Maintain turnaround times < 30 days for standard reviews and < 7 days for urgent safety signals.
Disclosure Completeness: 100% of participants must complete the cross-checked financial disclosure process via the National Medical Register prior to casting any evaluative vote.
Conflict Management: Zero unresolved major conflicts of interest. Any discovered omission results in immediate protocol referral.
Bypass Rate: Participant or institutional bypass rate must remain < 10%.
Conditions for Expansion: Expansion into Phase VI requires zero unmitigated critical safety events resulting from protocol failure, and the coalition's aggregate FCI must remain strictly below the 2500 (Highly Diversified) threshold.

4. Mandatory Operational Clauses & Readiness

Independent Evaluator Clause: The pilot’s evaluation shall be designed, approved, and reported by an evaluator independent of the Secretariat, participating Guild members, funders, deploying institutions, and model developers. The evaluator shall have authority to publish adverse findings without prior approval by any participating body.
Final Responsibility Clause: Every pilot activity shall identify a named legally accountable institution or person responsible for the resulting action. No CODE-AI output shall be interpreted as transferring, limiting, or extinguishing that responsibility.
Quorum and Voting Clause: No Council action shall be valid unless quorum (defined as at least two-thirds of appointed voting members), constituency participation, conflict exclusions, voting thresholds, dissent recording, and abstention rules are documented in advance.
Conflict-Verification Pathway: Conflicts must be independently verified via cross-reference with applicable public databases (e.g., the National Medical Register or Open Payments database) prior to seating, overriding self-reported declarations.
Data Ownership Rules: Raw evaluation data and pilot data remain the property of the originating institution, held under independent stewardship to guarantee operational portability and prevent vendor lock-in.

Appendix F — Version Change Log

v14.0 (Current): Executed comprehensive final review plan. Embedded operational clauses in Appendix E (Independent Evaluator, Final Responsibility, Quorum, Data Ownership). Added IRB/Ethics prerequisite to Roadmap. Added Change-Control Clause and Public Interest Access Clause to Appendix A. Fixed layout numbering, corrected FCI LaTeX thresholds to reflect true percentages, deployed the 4-tier Proportionality Decision Table, and enforced strict legal/epistemic wording globally.
v13.0: Embedded targeted structural edits addressing reviewer feedback. Applied required LaTeX syntax formatting to mathematical components. Formalized operational accountability tags within the Pilot Appendices. Clarified jurisdictional requirements for legal statements.
v12.0: Integrated strict independent research governance and pilot success thresholds. Replaced arbitrary scaling claims with empiric validation gates. Formalized the Funding Concentration Index (FCI) equation. Expanded core rules to include Pilot Suspension, Anti-Retaliation, and Data-Governance clauses. Deployed a standalone Pilot Protocol (Appendix E).

Works Cited

Numbered to match the in-text citation markers exactly. Where more than one URL appears under a single number, both sources support the same cited claim (a primary source paired with one corroborating or explanatory source); no number is used for a source that is not actually cited in the text above.

Empirical & Clinical Evidence

U.S. Food & Drug Administration guidance context — Chatterjee, N. et al., "Navigating US Regulation of Artificial Intelligence in Medicine," PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC11975648/
"Financial Conflicts of Interest in Clinical Practice Guidelines: A Systematic Review," PMC (37 studies; 14,764 authors; 45% FCOI; variation by specialty). https://pmc.ncbi.nlm.nih.gov/articles/PMC8105509/
"Financial Conflicts of Interest among U.S. Physician Authors of 2020 Clinical Practice Guidelines: A Cross-Sectional Study" (preprint, with subsequent journal publication). https://www.medrxiv.org/content/10.1101/2022.10.14.22281111.full | https://bmjopen.bmj.com/content/13/1/e069115
"Industry Payments to American Society of Hematology Clinical Practice Guideline Authors," Blood Advances. https://ashpublications.org/bloodadvances/article/8/13/3549/515579/Industry-payments-to-American-Society-of
National Medical Commission (India), Indian/National Medical Register. https://www.nmc.org.in/information-desk/indian-medical-register/

Institutional & Technical Precedents

World Wide Web Consortium, W3C Patent Policy (royalty-free licensing framework). https://www.w3.org/Consortium/Patent-Policy-20040205/
Apache Software Foundation, "A Primer on ASF Governance" (sponsorship separated from corporate governance). https://www.apache.org/foundation/governance/
IETF, RFC 7282 — "On Consensus and Humming in the IETF"; and "Supporting Participants," IETF Chairs Resources. https://datatracker.ietf.org/doc/rfc7282/ | https://chairs.ietf.org/en/participants

Legal & Regulatory Authorities

Internal Revenue Service, Exempt Organizations technical guidance on the public support test under IRC §§ 509(a)(1) and 170(b)(1)(A)(vi). https://www.irs.gov/pub/irs-tege/eotopicj93.pdf
American Society of Mechanical Engineers, Inc. v. Hydrolevel Corp., 456 U.S. 556 (1982). https://supreme.justia.com/cases/federal/us/456/556/ | https://en.wikipedia.org/wiki/American_Society_of_Mechanical_Engineers,_Inc._v._Hydrolevel_Corp.
North Carolina State Board of Dental Examiners v. FTC, 574 U.S. 494, 135 S. Ct. 1101 (2015); and FTC, "The When and What of Active Supervision." https://supreme.justia.com/cases/federal/us/574/494/ | https://harvardlawreview.org/print/vol-129/north-carolina-state-board-of-dental-examiners-v-ftc-2/ | https://www.ftc.gov/enforcement/competition-matters/2015/10/when-what-active-supervision

CODE-AI: Coalition of Domain Experts For AI