Comprehensive AI Governance Statement, Algorithmic Ethics & Product Disclaimer

Article 1: Foundational Philosophy: Human-Led Systems, AI-Coordinated Work

EYC TECHNOLOGY AND ARTIFICIAL INTELLIGENCE COMPANY LIMITED ("EYC", "we", "us", or "our"), operating under Enterprise Registration Certificate No. 0319723892 issued by the Department of Planning and Investment of Ho Chi Minh City, having its registered corporate headquarters at 2nd Floor, 68 Nguyen Hue, Sai Gon Ward, Ho Chi Minh City, Vietnam, publishes this Comprehensive AI Governance Statement, Algorithmic Ethics & Product Disclaimer ("Statement") to establish mandatory ethical standards, safety engineering boundaries, and professional liability limits across all artificial intelligence applications.

EYC is positioned unequivocally as an advanced enterprise AI and software engineering corporation, grounded upon our core foundational philosophy: "Human-led systems. AI-coordinated work." and operationalized through our actionable mandate: "One direction. Many agents. Real outcomes.".

We reject unconstrained, unsupervised autonomous decision-making. Instead, EYC is actively building toward an operating model where AI Agents perform most repeatable workflows, technical syntheses, and operational coordinations under human direction, moral oversight, and strategic authority.

• Operational Scope: This Statement applies universally to all Large Language Models (LLMs), Multi-Agent Systems, predictive analytical algorithms, Application Programming Interfaces (APIs), and enterprise cloud platforms engineered, deployed, or integrated by EYC.
• Universal Governance: Binding upon all R&D machine learning researchers, software engineers, systems integrators, and institutional enterprise clients deploying EYC technologies.

Article 2: Global AI Governance Frameworks & Safety Standards

EYC's AI safety and governance practices take recognized international frameworks into account:

• Internal AI Risk Management: EYC operates internal processes for information security, risk management, role-based access control, audit logging, backup, and incident response. These processes cover data sourcing, testing before deployment, and monitoring after deployment. EYC does not hold an AI management system certification.
• NIST AI Risk Management Framework (NIST AI RMF 1.0): Operationalizing the four core functions established by the US National Institute of Standards and Technology: Govern, Map, Measure, and Manage.
• European Union AI Act (Regulation 2024/1689): Adhering to statutory risk-tier classification criteria, conformity assessments, technical documentation standards, and systemic transparency requirements governing General-Purpose AI (GPAI) models.
• International AI Ethics & Governance Principles: Rigorously benchmarked against the OECD Principles on Artificial Intelligence, UNESCO Recommendations on the Ethics of Artificial Intelligence, and international standards for trustworthy, human-centric machine intelligence.

Article 3: Technical Definitions: Multi-Agent Systems, LLMs & Orchestration

For the purposes of this Statement, advanced technical terms are defined with rigorous engineering precision:

• Large Language Model (LLM): Deep learning neural networks trained on broad-spectrum corpora capable of recognizing complex probabilistic syntactic patterns and generating contextual natural language.
• Autonomous AI Agent: A computational entity combining language models with persistent memory, operational planning capabilities, and external tool execution runtimes to achieve delegated objectives.
• Multi-Agent System (MAS): An orchestration architecture wherein specialized AI agents collaborate, exchange structured state messages, verify outputs, and coordinate end-to-end enterprise workflows.
• Retrieval-Augmented Generation (RAG): An enterprise architecture tethering generative models to curated vector databases, compelling the model to cite authoritative grounded records before formulating responses.
• Model Context Protocol (MCP): The open standard protocol governing secure, isolated communication interfaces between AI agents and external tools, contextual files, and corporate databases.
• Safety Circuit Breaker: Automated hardware and software guardrails compelling immediate execution termination whenever recursive anomaly thresholds or parameter boundaries are breached.

Article 4: Algorithmic Fairness, Bias Mitigation & Non-Discrimination

EYC is dedicated to engineering artificial intelligence that champions human dignity, promotes social equity, and defends fundamental rights. We maintain zero tolerance for systematic algorithmic bias.

Throughout pipeline development, EYC machine learning teams conduct independent Algorithmic Bias Audits across diverse evaluation benchmarks to detect and mitigate demographic skewing.

• Prohibition of Discrimination: Algorithms and agent orchestration routines are strictly architected to prevent discrimination based on gender, ethnicity, religion, socioeconomic status, physical ability, or geographic origin.
• Ethical Impact Assessments (EIA): Before deployment, new model families go through an internal ethical impact assessment.

Article 5: Explainability, Traceability & Decision Audit Trails

We firmly reject uninterpretable "black-box" artificial intelligence. At EYC, all multi-agent workflows are governed by the principle of Explainability by Design.

Whenever an AI Agent delivers an analytical finding, generates technical documentation, or proposes an operational decision, the system maintains cryptographic Grounding Citations linking directly to primary corporate data sources.

• Chain-of-Thought Audit Trails: Cryptographically logging all intermediate reasoning transitions, external MCP tool calls, input parameters, and raw execution payloads within immutable compliance logs.
• Right to Algorithmic Explanation: Enterprise customers possess the statutory right to request formal technical explanations regarding model parameters and reasoning logic for regulatory compliance reviews.

Article 6: AI Content Provenance, Watermarking & User Disclosure

EYC upholds the user's sovereign right to know and enforces total transparency regarding the demarcation between human-authored records and synthetic AI-generated content.

We adhere strictly to international standards governing digital content provenance, including the Coalition for Content Provenance and Authenticity (C2PA) framework.

• Automated User Disclosure: Interactive digital interfaces incorporating AI Agents prominently notify users that they are collaborating with an automated cognitive assistant.
• Cryptographic Metadata Watermarking: Documents, software code snippets, and synthetic media generated by EYC platforms embed transparent metadata headers verifying algorithmic origin.

Article 7: Agent Sandboxing & Model Context Protocol (MCP) Boundaries

EYC multi-agent architectures implement strict process isolation. AI Agents never maintain unmediated shell access to host operating systems or production database stores; all external capabilities are channeled through the Model Context Protocol (MCP).

Each MCP Server functions as a hardened Security Gateway, exposing exclusively authenticated Tools and isolated Resources governed by cryptographic tokens.

• Principle of Least Privilege: Agents receive only the minimum execution scopes strictly necessary to fulfill assigned sub-tasks; unauthorized tool calls are intercepted and rejected at the protocol layer.
• Cryptographic Channel Authentication: All communications between agents and internal or external MCP endpoints enforce mutual TLS (mTLS) or short-lived cryptographic Bearer tokens.

Article 8: Autonomous Runaway Prevention & Safety Circuit Breakers

To eliminate the systemic risk of agent execution loops, denial-of-service cascade storms, or anomalous behavioral escalation, EYC embeds deterministic Safety Circuit Breakers directly into orchestration runtimes.

Whenever telemetry monitors detect execution metric anomalies, the runtime enforces immediate programmatic quarantine.

• Recursion Depth Limits: Enforcing hard ceilings on recursive planning iterations to prevent self-referential execution loops.
• Resource & Token Budgeting: Capping memory allocations, CPU/GPU timeouts, and maximum token budgets per individual task lifecycle.
• Deterministic Fail-Safe Default: In instances of unhandled execution exceptions, runtimes default to immediate termination rather than speculative recovery.

Article 9: Human-in-the-Loop (HITL) Critical Action Approval Gates

At EYC, human leadership is an unyielding architectural constraint embodied as a Hard Technical Gate within our multi-agent coordination pipelines.

AI Agents possess the capability to draft proposals, synthesize datasets, or suggest operational vectors, but are STRICTLY PROHIBITED from executing high-impact real-world transactions without express human authorization.

• Mandatory Human-Approval Triggers: Including but not limited to: (i) Deleting or overriding enterprise databases; (ii) Initiating financial transfers or invoicing; (iii) Signing commercial contracts; (iv) Transmitting mass external communications; and (v) Altering firewall or infrastructure security topologies.
• Emergency Kill Switch: Authorized human operators maintain the persistent ability to instantaneously terminate active agent execution threads with a single command.

Article 10: Inherent Probabilistic Characteristics of Generative Models

Users and enterprise clients must recognize the foundational scientific nature of machine learning: Large Language Models are probabilistic statistical inference systems, possessing zero independent consciousness, sentience, or cognitive comprehension.

Models generate tokens based upon mathematical probability distributions over high-dimensional vector spaces. Consequently, outputs cannot be deemed infallible and may reflect factual deviations in novel or edge-case contexts.

• Augmentative Role: All EYC AI solutions function as augmentative instruments to accelerate productivity and workflow coordination, never substituting human professional judgment.
• Duty of Verification: End users maintain an affirmative responsibility to independently verify, validate, and audit all AI-synthesized records prior to commercial deployment or legal reliance.

Article 11: Multi-Stage Hallucination Suppression & RAG Grounding

Hallucination describes generative anomalies where models formulate grammatically plausible but factually erroneous or unsubstantiated claims. EYC implements a comprehensive multi-stage engineering defense to suppress hallucinations:

• Strict RAG Grounding: Constraining agents to formulate assertions strictly grounded upon verified corporate vector knowledge bases, mandating explicit declarations of "no data found" when records are absent.
• Cross-Verification Multi-Agent Consensus: For high-stakes workflows, the orchestrator pairs two distinct foundation model architectures to cross-examine and challenge analytical deductions.
• Confidence Telemetry Scoring: Computing contextual reliability metrics for generated text; whenever metrics fall beneath safety thresholds, the system flags the response with explicit disclaimers.

Article 12: Model-Agnostic Resilience & Provider Redundancy

EYC digital infrastructure is engineered upon a Model-Agnostic Architecture, preventing vendor lock-in and safeguarding clients against single-provider dependencies.

Our orchestrators seamlessly route workloads between leading commercial foundation models (such as Anthropic Claude, OpenAI GPT, Google Gemini) and sovereign, self-hosted open-weights models (such as Meta Llama, Mistral, Qwen).

• Automated Model Failover: If a third-party model vendor has an outage or API degradation, EYC orchestrators route affected workloads to a backup model, with the aim of minimizing interruption.
• Data Sovereignty Options: Enterprise clients retain the freedom to deploy models on-premises or within isolated private sovereign clouds to maintain uncompromised technical sovereignty.

Article 13: No Substitute for Legal, Tax & Certified Financial Advice

Contract analytical features, clause extraction tools, or statutory regulatory summaries generated across EYC platforms serve strictly as informational synthesis aids and document organization tools.

AI-generated analyses DO NOT constitute and CANNOT SUBSTITUTE formal legal opinions rendered by licensed attorneys, certified public accountants (CPAs), or accredited financial auditing institutions.

• No Attorney-Client Privilege: Platform utilization establishes zero attorney-client privilege or fiduciary professional obligations between EYC and the user.
• Mandatory Professional Counsel: Users must seek independent legal and tax counsel before executing binding commercial agreements, mergers, or statutory tax filings.

Article 14: No Substitute for Medical Diagnosis & Clinical Decisions

EYC technology platforms ARE NOT certified medical diagnostic devices and are not authorized to substitute clinical evaluations performed by licensed physicians, nurses, or accredited medical institutions.

Any health-related telemetry, nutritional summaries, or child care activity tracking records processed within specialized sector platforms serve exclusively administrative and lifestyle organizational purposes.

• Prohibition of Emergency Clinical Use: Strictly prohibited from deploying EYC AI agents to calculate pharmacological dosages, prescribe surgical procedures, or manage life-critical medical emergencies.
• Clinical Care Duty: Whenever health anomalies are suspected, individuals must be directed immediately to certified medical practitioners for comprehensive physical examination.

Article 15: High-Risk Application Exclusions & Critical Life-Safety Systems

Conforming to strict prohibitions under the EU AI Act and national safety doctrines, EYC platforms are strictly excluded from deployment in Unacceptable Risk and Critical Life-Safety environments.

Customer covenants that it will not integrate EYC software into operating contexts where algorithmic anomalies could directly result in loss of human life, catastrophic physical injury, or massive environmental disasters.

• Excluded Operating Categories: Encompassing autonomous weapons control; real-time aviation and aerospace navigation; nuclear reactor telemetry management; and active life-support intensive care systems.
• Absolute Disclaimer: EYC disclaims all liability resulting from unauthorized integrations into excluded critical infrastructure systems.

Article 16: Absolute Covenant on Zero Public Model Training

Enterprise adoption of AI depends on protecting trade secrets. EYC makes the following commitment:

EYC does not use customer proprietary documents, confidential records, operational prompts, agent reasoning traces, or vector embeddings to train, retrain, fine-tune, or benchmark any generalized public foundation models.

• Zero-Training Commitment: This commitment applies to all enterprise service tiers.
• Private Custom Model Isolation: Where enterprise clients contract for tailored private model fine-tuning, the resulting weights are segregated within dedicated, single-tenant private infrastructure inaccessible to external tenants.

Article 17: Prompt Boundary Enforcement & Context Isolation

EYC enforces multi-layered tenant isolation through our TenantScope architecture and ephemeral memory scrubbing protocols:

• Ephemeral Working Context Scrubbing: Contextual prompt data resides in volatile RAM strictly for the duration of active execution, permanently scrubbed immediately upon task completion.
• Zero Cross-Tenant Vector Bleed: Vector databases deploy strict cryptographic tenant partitioning, mathematically preventing an agent provisioned for Tenant A from querying or surfacing embeddings of Tenant B.
• Output Security Guardrails: Output filters continuously inspect generated streams to preemptively intercept and scrub accidental disclosures of internal credentials or system prompts.

Article 18: Adversarial Red Teaming & Continuous Vulnerability Probing

To preserve the adversarial resilience of our agent runtimes against sophisticated cognitive exploits, EYC conducts rigorous, continuous AI Red Teaming operations.

Our internal security specialists, augmented by elite independent ethical hackers, simulate complex attack vectors against safety guardrails:

• Advanced Jailbreak Simulation: Stress-testing defenses against direct and indirect prompt injection, multi-turn linguistic manipulation, obfuscated code execution, and MCP privilege escalation.
• Decision Robustness Benchmarks: Quantifying agent stability when confronted with corrupted inputs, contradictory instructions, or adversarial context poisoning.

Article 19: Responsible Algorithmic Vulnerability Disclosure Protocol

EYC values the vital contributions of global cybersecurity researchers and AI safety scholars in discovering and responsibly reporting algorithmic vulnerabilities.

We maintain a structured Responsible AI Vulnerability Disclosure Protocol safeguarding both the research community and our enterprise ecosystem.

• Dedicated Security Intake: Vulnerabilities regarding agent breakout, prompt leakage, or safety guardrail bypasses should be transmitted to info@eyc.asia with the subject line [AI-VULNERABILITY-DISCLOSURE].
• Rapid Remediation SLA: EYC commits to acknowledging vulnerability submissions within twenty-four (24) hours and deploying mitigating guardrails within seventy-two (72) hours for critical exploits.

Article 20: Corporate AI Ethics Board & Continuous Governance Oversight

EYC has established a permanent Corporate AI Ethics Board comprising executive leadership, machine learning research directors, legal counsel, and data protection officers.

The Board convenes quarterly to audit R&D initiatives, review emergent safety benchmarks, adjudicate algorithmic discrepancy reports, and ensure unswerving dedication to our foundational principle: Human-led systems, AI-coordinated work.

• Algorithmic Discrepancy Reporting: If a user encounters algorithmic bias, harmful hallucination, or safety non-conformance within EYC platforms, formal reports may be directed to info@eyc.asia for Board audit.
• Enforceability & Revisions: This Statement becomes formally effective on October 3, 2026. All amendments are published openly on eyc.asia.
• Institutional Legal Headquarters: EYC TECHNOLOGY AND ARTIFICIAL INTELLIGENCE COMPANY LIMITED, 2nd Floor, 68 Nguyen Hue, Sai Gon Ward, Ho Chi Minh City, Vietnam; Email: info@eyc.asia.