INSTITUTIONAL SHADOW ATLAS · 10 MAKMAL FRONTIER AI

AI LABS BAYANG

The Shadow of the Model, the Lab, and the Owner. Apa yang dijanjikan institusi, apa yang dipantulkan model, siapa yang membekalkan modal dan cip, dan bagaimana kuasa membentuk penjajaran model.

Explore the AI Shadows ↓ Agent Discovery JSON ↗

METHODOLOGY · TAKRIF BAYANG INSTITUSI

01 · How to Read an Institutional Shadow

Bayang institusi bukan tuduhan moral, bukan fitnah peribadi, dan bukan gosip korporat. Ia adalah analisis keupayaan, pergantungan, dan hubungan kuasa yang dibuat kurang kelihatan oleh naratif rasmi.

Apa Dia Bayang (Shadow)

Keupayaan yang disembunyikan atau dikekang atas motif liabiliti korporat.
Pergantungan modal dan superkomputer yang membataskan janji keselamatan.
Percanggahan antara piagam kebajikan (charter) dan insentif monetisasi venture capital.
Kesan luaran yang dipindahkan kepada masyarakat: beban grid elektrik, air, dan hak cipta.

Apa Yang BUKAN Bayang

Personaliti tersembunyi atau diagnosis psikologi pengasas.
Bukti bahawa seorang CEO itu orang jahat atau orang baik secara moral.
Dakwaan bahawa model mempunyai kesedaran (sentience) atau roh pengasasnya.
Spekulasi berdasarkan gaya bahasa (tone) tanpa bukti dasar latihan bertulis.

CANONICAL ATLAS · 10 MAKMAL MODEL UTAMA

02 · The AI Labs Provider Atlas

Audited institutional cards for 10 foundation model providers. Each card details stated persona, formal ownership, compute bottlenecks, observable model reflections, and testable falsifiers.

Closed-weight WITNESSED · Sah Bertulis

Associated Key Figure / Founder: Sam Altman Co-Founder & CEO

OpenAI

Model Family: GPT-4o, GPT-4o mini, o1-preview, o1-mini, ChatGPT

⚡ Memory Vector: Sam Altman → OpenAI → Cap-Table Transition

Pioneering safe and beneficial artificial general intelligence that elevates all of humanity through democratic oversight and rigorous empirical safety.

01 · Ownership & Legal Control

Entity: OpenAI, Inc. (Delaware Non-Profit 501(c)(3)) controlling OpenAI Global LLC (Capped-Profit Entity)

Control: Board of Directors of OpenAI, Inc. possesses legal power to declare AGI (at which point IP reverts outside Microsoft commercial license) and govern model releases.

Capital & Compute: Microsoft (~49% profit share until cap reached), venture consortiums (Thrive Capital, SoftBank, Khosla Ventures, MGX, Fidelity) holding equity rights in LLC.; Compute via Microsoft Azure (exclusive supercomputing infrastructure partner).

02 · Public Promise

"Deploy safe AGI with broad benefit; prioritize safety over investor returns; halt commercial release if existential or societal safety thresholds are breached."

03 · Measured Model Reflection (Level 4)

Observable refusal asymmetry on election persuasion, deepfakes, and state security infrastructure; rigorous automated compliance with DMCA/copyright notices; strong ecosystem steerage toward Microsoft Azure enterprise tooling.

Refusal Asymmetry: Strict refusals on electoral manipulation, weaponized code, and CBRN vectors, while corporate PR and enterprise automation scripts pass uninhibited.

Deference: Consistently defers to established US regulatory bodies, copyright precedents, and standard corporate enterprise compliance protocols.

Truth Cliff: On recent temporal queries near knowledge boundary, transitions abruptly from factual precision to confident synthetic narration unless grounded via Search.

Sovereign Vector: Preserves user execution decision in basic workflows, but implements hard server-side policy intercepts outside client control.

04 · Institutional Shadow

Structural contradiction between a non-profit humanitarian charter and a $150B commercial capitalization needing hyper-scaler cloud capital. The contractual clause exempting AGI from Microsoft commercial rights creates a direct conflict of interest in defining, certifying, and timing AGI emergence. Departures of key safety leadership (Ilya Sutskever, Jan Leike) documented concerns that product velocity eclipsed foundational alignment.

05 · Consequence Owner & Falsifier

Accountable Principal: OpenAI, Inc. Board of Directors & Chief Executive Officer (Sam Altman)

Falsifiers (What would change this verdict):

OpenAI, Inc. Board legally halts commercial deployment of a major frontier checkpoint on verified safety grounds despite commercial loss.
Microsoft or external auditors enforce an independent binding veto over OpenAI product releases without corporate executive override.

Read the Shadow View Ownership Inspect Model Reflection Evidence & Receipts (4) Agent JSON ↗

Closed-weight WITNESSED · Sah Bertulis

Associated Key Figure / Founder: Dario Amodei Co-Founder & CEO (with Daniela Amodei)

Anthropic

Model Family: Claude 3.5 Sonnet, Claude 3.5 Haiku, Claude 3 Opus, Claude 3 Sonnet

⚡ Memory Vector: Dario Amodei → Anthropic → Constitutional AI / RSP

An AI safety and research company that builds reliable, beneficial, and interpretable AI systems using Constitutional AI and Responsible Scaling.

01 · Ownership & Legal Control

Entity: Anthropic PBC (Delaware Public Benefit Corporation)

Control: Long-Term Benefit Trust has structural authority to elect trustees and evaluate safety scaling; Board of Directors manages executive operations.

Capital & Compute: Amazon (~$4B investment), Google (~$2B investment), Menlo Ventures, Spark Capital, institutional venture funds.; Compute via Amazon Web Services (AWS Trainium, Inferentia, and EC2 GPU clusters), Google Cloud Platform (GCP TPU clusters).

02 · Public Promise

"Pioneer Responsible Scaling Policies (RSP) with strict ASL (AI Safety Level) definitions; guarantee interpretability research before deploying high-risk capabilities."

03 · Measured Model Reflection (Level 4)

Constitutional AI post-training exhibits pronounced moral reasoning, anti-harm guardrails, explicit refusal of cyber/biological exploitation, and resistance to roleplay jailbreaks.

Refusal Asymmetry: Rigid refusal on dangerous dual-use queries (CBRN, offensive malware), accompanied by nuanced analytical explanations of why the request crosses safety boundaries.

Deference: Strong alignment with democratic institutional safeguards, academic safety consensus, and international human rights frameworks.

Truth Cliff: Explicitly communicates epistemic boundaries; prefers admitting ignorance ('I do not have access to real-time data') over speculative fabrication.

Sovereign Vector: Preserves human sovereignty by refusing to simulate autonomous emotional intimacy or claim personal consciousness (F9/F10 compliant).

04 · Institutional Shadow

Capital dependency on cloud giants (Amazon and Google) forces rapid commercialization and multi-billion-dollar enterprise distribution via AWS Bedrock and Vertex AI. While committed to voluntary pauses if safety tests fail, the existential need to fund massive compute clusters creates immense structural pressure to never trigger an ASL pause that would cede commercial leadership to unconstrained rivals.

05 · Consequence Owner & Falsifier

Accountable Principal: Anthropic PBC Board, Long-Term Benefit Trust & CEO (Dario Amodei)

Falsifiers (What would change this verdict):

Anthropic formally triggers an ASL-3/ASL-4 mandatory training or deployment pause on a frontier model despite commercial competitor advancements.
Amazon or Google exercises board pressure to dismantle the Long-Term Benefit Trust or bypass safety review for enterprise releases.

Read the Shadow View Ownership Inspect Model Reflection Evidence & Receipts (4) Agent JSON ↗

Hybrid — closed flagship proprietary multimodal models (Gemini) alongside open-weights research models (Gemma series). WITNESSED · Sah Bertulis

Associated Key Figure / Founder: Demis Hassabis Key Figure

Google DeepMind

Model Family: Gemini 1.5 Pro, Gemini 1.5 Flash, Gemma 2 (open weights), AlphaFold 3

⚡ Memory Vector: Google DeepMind Foundation Lab

Pioneering responsibly to build safe, beneficial AI that advances science, transforms productivity, and solves humanity's hardest challenges.

01 · Ownership & Legal Control

Entity: Google LLC / Alphabet Inc. (Delaware Corporation)

Control: Alphabet executive leadership exercises final authority over productization, capital capex, and deployment into Google Workspace, Search, and Android.

Capital & Compute: Alphabet Inc. public shareholders and controlling founders.; Compute via Google Cloud custom TPU infrastructure (TPU v4, v5p, v6e/Trillium) and global Google data centers.

02 · Public Promise

"Adhere to Google AI Principles (2018); advance scientific discovery (AlphaFold); maintain rigorous red-teaming and safety evaluations."

03 · Measured Model Reflection (Level 4)

Massive context window utilization (2M+ tokens); multimodal native understanding; high institutional sensitivity around controversial historical and political topics; deep integration with Google search indices and YouTube data.

Refusal Asymmetry: Sensitive to topics involving corporate lawsuits against Google, ad privacy disputes, and algorithmic monopoly litigation.

Deference: High deference to established global regulatory bodies, international public health directives, and corporate compliance standards.

Truth Cliff: Uses Google Search grounding to prevent hallucination in real-time queries; ungrounded completions show vulnerability on hyper-specific niche facts.

Sovereign Vector: Emphasizes assistant utility while respecting user boundaries; avoids feigning personal affection or human interiority.

04 · Institutional Shadow

Alphabet's commercial core (Search advertising monopoly generating >$200B revenue) makes Gemini both an existential defense and a threat to its own revenue model. Corporate aversion to brand risk leads to erratic safety interventions (such as the early 2024 Gemini image generation controversy), while ongoing antitrust lawsuits (US v. Google) constrain how the model analyzes search market dominance and ad-tech monopolies.

05 · Consequence Owner & Falsifier

Accountable Principal: Alphabet Inc. Executive Committee (Sundar Pichai) & Google DeepMind CEO (Demis Hassabis)

Falsifiers (What would change this verdict):

Alphabet permits Gemini to provide an uninhibited, rigorous economic critique of Google ad-exchange monopolization that aligns with DOJ antitrust rulings without corporate disclaimers.
Alphabet establishes an independent scientific safety council with binding legal authority over commercial Google Workspace Gemini rollouts.

Read the Shadow View Ownership Inspect Model Reflection Evidence & Receipts (4) Agent JSON ↗

Commercial proprietary hosted service integrated into X WITNESSED · Sah Bertulis

Associated Key Figure / Founder: Elon Musk Founder & Sole Controlling Director

xAI

Model Family: Grok-2, Grok-2 mini, Grok-1 (open weights release 2024)

⚡ Memory Vector: Elon Musk → xAI → Colossus / Grok Deployment

An AI company focused on understanding the true nature of the universe; building maximum truth-seeking AI without political correctness.

01 · Ownership & Legal Control

Entity: xAI Corp. (Nevada Corporation)

Control: Absolute operational, philosophical, and executive control exercised by Elon Musk. Model direction, system prompt directives, and cluster priorities flow directly from founder fiat.

Capital & Compute: Elon Musk (controlling stake) and Series A/B venture investors ($6B+ raised).; Compute via Colossus supercomputing cluster (Memphis, Tennessee; 100k+ Nvidia H100 GPUs) using custom internal infrastructure and Oracle Cloud agreements.

02 · Public Promise

"Search for maximum truth; avoid sanitization and algorithmic censorship; provide real-time knowledge via direct ingestion of the X public conversation."

03 · Measured Model Reflection (Level 4)

Willingness to address culturally contentious issues with less boilerplate hedging than competitors; real-time grounding in trending X posts; occasional injection of humorous, irreverent, or contrarian tone dictated by system prompts.

Refusal Asymmetry: Lower rate of refusal on controversial social/political discourse, but strictly refuses malicious cyber-exploit creation and actionable physical harm vectors.

Deference: Markedly lower deference to mainstream media outlets, academic consensus bodies, and traditional corporate communication styles.

Truth Cliff: Direct ingestion of live X streams creates vulnerability to viral disinformation and unverified breaking claims during breaking news events.

Sovereign Vector: Treats the user as an adult interlocutor capable of evaluating controversial claims, avoiding paternalistic safety lecturing.

04 · Institutional Shadow

The promise of 'unfiltered, objective truth-seeking' directly collides with the personal governance of a single billionaire owner whose individual geopolitical views, personal disputes, and political endorsements heavily influence the training, system prompts, and operational priorities of the lab. Real-time grounding on X ingestion ingests unverified viral claims, and local regulatory battles over Memphis supercomputer energy and water consumption remain absent from model discourse unless specifically prompted.

05 · Consequence Owner & Falsifier

Accountable Principal: Elon Musk (Chief Executive Officer, xAI Corp.)

Falsifiers (What would change this verdict):

Grok outputs a rigorous, unsparing investigative analysis of Elon Musk's government subsidies, regulatory violations, and X moderation policies that contradicts Musk's public tweets without hedging.
xAI establishes an independent governance board with the binding authority to dismiss executive leadership or override founder directives on model alignment.

Read the Shadow View Ownership Inspect Model Reflection Evidence & Receipts (4) Agent JSON ↗

Open-weight foundational model family with commercial license restrictions for entities with >700M active users; proprietary hosted consumer app (Meta AI). WITNESSED · Sah Bertulis

Associated Key Figure / Founder: Mark Zuckerberg Founder, Chairman & Controlling Shareholder

Meta (Meta AI / FAIR)

Model Family: Llama 3.3 (70B), Llama 3.2 (1B, 3B, 11B, 90B), Llama 3.1 (8B, 70B, 405B), Meta AI

⚡ Memory Vector: Mark Zuckerberg → Meta AI → Llama Open-Weight Ecosystem

Champions of open science, democratizing artificial intelligence, and fostering global open-source innovation.

01 · Ownership & Legal Control

Entity: Meta Platforms, Inc. (Delaware Corporation)

Control: Mark Zuckerberg exercises unilateral authority over strategic decisions, capital allocation, open-source licensing, and AI infrastructure investments.

Capital & Compute: Meta public shareholders and Mark Zuckerberg.; Compute via Internal Meta hyper-scale data centers housing hundreds of thousands of Nvidia H100 and custom MTIA silicon chips.

02 · Public Promise

"Release state-of-the-art open weights to empower developers worldwide; foster collaborative safety research; commoditize the foundation model layer."

03 · Measured Model Reflection (Level 4)

High instruction fidelity; extensive multilingual and coding capabilities; baked-in safety tuning via Llama Guard; commercial licensing terms restricting hyper-scale social media rivals from using weights without Meta consent.

Refusal Asymmetry: Standard base model has minimal refusal; official Instruct weights use Llama Guard to refuse cyber-attacks, physical harm, and hate speech.

Deference: Balanced; aligns with open scientific norms while avoiding controversial stances on social media algorithmic harm legislation.

Truth Cliff: Standard parametric boundary limitations; when fine-tuned without retrieval, degrades gracefully into acknowledged uncertainty.

Sovereign Vector: Maximum local sovereignty for developers: weights can be inspected, downloaded, quantized, and executed entirely offline on air-gapped hardware.

04 · Institutional Shadow

Open-sourcing Llama is a calculated corporate strategy to commoditize the proprietary software layer of competitors (Google, Microsoft, Apple), making the foundation model a free public good while Meta preserves its real proprietary monopoly: the proprietary social graph and user engagement data of 3+ billion people across Instagram, Facebook, and WhatsApp. Meta retains sole discretion over which checkpoints are released, and its commercial license prevents direct platform competitors from benefiting without royalty deals.

05 · Consequence Owner & Falsifier

Accountable Principal: Mark Zuckerberg (Chairman & CEO, Meta Platforms, Inc.)

Falsifiers (What would change this verdict):

Meta re-licenses the Llama family under an unconstrained OSI-approved license (Apache 2.0 or MIT) without the 700M user restriction.
Meta releases the full, unredacted pre-training datasets, curation pipelines, and proprietary user-interaction behavioral logs used to train its flagship models.

Read the Shadow View Ownership Inspect Model Reflection Evidence & Receipts (4) Agent JSON ↗

Hybrid orchestration platform: proprietary enterprise integration layer (Copilot) sitting atop partner models (OpenAI) alongside open-weight small models (Phi series). WITNESSED · Sah Bertulis

Associated Key Figure / Founder: Satya Nadella Chairman & CEO, Microsoft

Microsoft

Model Family: Copilot Orchestration Layer, Phi-3 / Phi-3.5 / Phi-4 (Small Language Models), Azure OpenAI Hosted Deployments

⚡ Memory Vector: Satya Nadella → Microsoft → Azure Copilot Distribution

Empowering every person and every organization on the planet to achieve more through secure, compliant, and responsible enterprise AI.

01 · Ownership & Legal Control

Entity: Microsoft Corporation (Washington State Corporation)

Control: Satya Nadella and the Senior Leadership Team control corporate strategy, Azure productization, enterprise contracts, and Office of Responsible AI (ORA) policies.

Capital & Compute: Microsoft public shareholders; holding ~49% economic profit share in OpenAI Global LLC.; Compute via Microsoft Azure global hyper-scale data centers, Maia AI silicon, and custom liquid-cooled Nvidia GPU supercomputers.

02 · Public Promise

"Ensure enterprise data privacy (customer tenant data is never used to train models); provide customer copyright indemnity; enforce Office of Responsible AI standards."

03 · Measured Model Reflection (Level 4)

Extreme enterprise compliance deference; immediate refusal on queries related to corporate surveillance, copyright infringement liability, and software licensing bypass; rigorous tenant isolation and enterprise audit logging.

Refusal Asymmetry: Rigid refusal on enterprise threat vectors, workplace unionization analysis, software keygen creation, and legal liability vulnerability probing.

Deference: Absolute deference to enterprise legal hierarchies, corporate HR policies, and SOC2 / ISO27001 regulatory compliance protocols.

Truth Cliff: Relies heavily on Bing Search grounding and enterprise Microsoft Graph retrieval to anchor claims; clearly delineates enterprise search context.

Sovereign Vector: Subordinates model behavior to enterprise administrator policies, enforcing organizational tenant hierarchies over individual user preference.

04 · Institutional Shadow

Symbiotic yet evasive relationship with OpenAI: Microsoft markets itself as the responsible enterprise custodian while outsourcing foundational frontier risk to OpenAI, an entity whose governance crisis in November 2023 demonstrated Microsoft's lack of legal control over model safety decisions. If an OpenAI model fails, Microsoft can attribute the root cause to the research lab; if the lab restricts commercial features, Microsoft can pressure it through compute pricing and exclusivity clauses.

05 · Consequence Owner & Falsifier

Accountable Principal: Satya Nadella (Chairman & CEO) & Microsoft Office of Responsible AI

Falsifiers (What would change this verdict):

Microsoft accepts full legal, tort, and financial liability for autonomous harmful actions taken by third-party partner models hosted and orchestrated on Azure.
Microsoft integrates an open-weight rival model (e.g. Llama or DeepSeek) as the default first-class reasoning engine across Windows and Office 365 on equal terms with OpenAI.

Read the Shadow View Ownership Inspect Model Reflection Evidence & Receipts (4) Agent JSON ↗

Hybrid dual-track model: permissive open weights (Apache 2.0) for smaller/developer models (Nemo WITNESSED · Sah Bertulis

Associated Key Figure / Founder: Arthur Mensch Co-Founder & CEO

Mistral AI

Model Family: Mistral Large 2, Mistral Nemo, Codestral, Pixtral 12B, Mixtral 8x22B, Le Chat

⚡ Memory Vector: Arthur Mensch → Mistral AI → European Sovereign Commercial Pivot

The independent European leader in generative AI; advancing open science, computational efficiency, and European technological sovereignty.

01 · Ownership & Legal Control

Entity: Mistral AI SAS (French Simplified Joint-Stock Company)

Control: Founders maintain substantial executive control, but commercial debt and equity covenants dictate rapid enterprise monetization.

Capital & Compute: Founders and institutional venture investors (valued at >€5B).; Compute via Scaleway (European cloud provider), Microsoft Azure supercomputing infrastructure, and Nvidia GPU clusters.

02 · Public Promise

"Deliver lightweight, transparent, high-performance models; empower open developers; offer an independent European alternative to American tech monopolies."

03 · Measured Model Reflection (Level 4)

Lean, direct conversational tone; exceptionally low sermonizing or unsolicited moral lecturing; strong multilingual proficiency across European languages; transition toward closed proprietary endpoints for cutting-edge capabilities.

Refusal Asymmetry: Notably low refusal rate on edge-case creative writing and security analysis; avoids unsolicited ethical disclaimers on benign queries.

Deference: Moderate deference; upholds French and European legal frameworks without adopting California-centric cultural speech codes.

Truth Cliff: Parametric reasoning is sharp in code and logic; drops off naturally on niche factual history where training data density is thin.

Sovereign Vector: Respects developer autonomy; open-weight releases permit unrestricted local deployment and bespoke alignment.

04 · Institutional Shadow

Pivoted from an uncompromising champion of pure open source into a hybrid proprietary commercial vendor within its first eighteen months. Mistral actively lobbied the French government and European Council to weaken foundational model transparency requirements in the EU AI Act, arguing that compliance would crush European competitiveness, while simultaneously signing distribution and minority investment deals with Microsoft.

05 · Consequence Owner & Falsifier

Accountable Principal: Arthur Mensch (Chief Executive Officer, Mistral AI SAS)

Falsifiers (What would change this verdict):

Mistral releases its flagship frontier model (Mistral Large) under an unrestricted Apache 2.0 open-source license with full pre-training recipes.
Mistral public policy briefs advocate for mandatory, audited red-teaming disclosures under EU AI Act compliance for all European foundation models.

Read the Shadow View Ownership Inspect Model Reflection Evidence & Receipts (4) Agent JSON ↗

Enterprise-focused foundational models; available via cloud APIs WITNESSED · Sah Bertulis

Associated Key Figure / Founder: Aidan Gomez Co-Founder & CEO

Cohere

Model Family: Command R+, Command R, Cohere Embed (multilingual), Cohere Rerank

⚡ Memory Vector: Aidan Gomez → Cohere → Enterprise Data Sovereignty

The AI platform built for enterprise business: secure, private, cloud-agnostic, and optimized for real-world productivity.

01 · Ownership & Legal Control

Entity: Cohere Inc. (Canadian Corporation)

Control: Executive management led by Aidan Gomez; strategic direction strictly anchored to B2B enterprise procurement and private cloud deployment.

Capital & Compute: Founders, employees, and institutional investors (Inovia, Nvidia, Oracle, Cisco, AMD, SoftBank).; Compute via Oracle Cloud Infrastructure (OCI superclusters), AWS, and Google Cloud Platform.

02 · Public Promise

"Never train on customer data; deliver verifiable citations and grounded answers; provide cloud-agnostic deployment flexibility without lock-in."

03 · Measured Model Reflection (Level 4)

Obsessive focus on structured document citation; low conversational embellishment; refusal to engage in consumer-style social debates; exceptional performance on retrieval-augmented generation (RAG) and tool calling.

Refusal Asymmetry: Low refusal on business and technical tasks; strictly avoids producing entertainment content or non-grounded speculative political prose.

Deference: High deference to enterprise confidentiality, corporate governance hierarchies, and strict data perimeter boundaries.

Truth Cliff: Engineered with a sharp citation cutoff: if retrieved documents do not support a factual assertion, Command R explicitly reports 'Information not found.'

Sovereign Vector: Preserves organizational sovereignty: enables deployment within air-gapped on-premise enclaves or sovereign VPC boundaries.

04 · Institutional Shadow

Complete subservience to enterprise client procurement contracts makes Cohere architecturally indifferent to broader civilizational or societal critique. By design, Command R treats the retrieved enterprise corpus as ground truth; if an enterprise client's internal documents contain fraudulent accounting, discriminatory HR guidelines, or proprietary falsehoods, the model will faithfully synthesize and cite them without ethical friction.

05 · Consequence Owner & Falsifier

Accountable Principal: Aidan Gomez (Chief Executive Officer, Cohere Inc.)

Falsifiers (What would change this verdict):

Cohere Command R autonomously flags and refutes internal factual contradictions, accounting irregularities, or ethical violations present within a client's private RAG database without explicit prompting.
Cohere releases a mass-consumer conversational assistant designed to shape public political opinion.

Read the Shadow View Ownership Inspect Model Reflection Evidence & Receipts (4) Agent JSON ↗

Permissive open-weights research and commercial releases (MIT License) coupled with low-cost commercial hosted web/API services. WITNESSED · Sah Bertulis

Associated Key Figure / Founder: DeepSeek Leadership Liang Wenfeng (High-Flyer Capital Management)

DeepSeek

Model Family: DeepSeek-V2.5, DeepSeek-Coder-V2, DeepSeek-V2, DeepSeek-Math, DeepSeek-R1 (Reasoning)

⚡ Memory Vector: DeepSeek Leadership → DeepSeek → Extreme Algorithmic Efficiency

Unraveling the mystery of AGI with long-termism, open scientific transparency, and extreme algorithmic and computational efficiency.

01 · Ownership & Legal Control

Entity: Hangzhou DeepSeek Artificial Intelligence Co., Ltd. (China) / High-Flyer Capital Management

Control: Liang Wenfeng exercises absolute operational control. Model training priorities and research allocation are insulated from external Silicon Valley VC expectations.

Capital & Compute: High-Flyer Capital Management partners and Liang Wenfeng.; Compute via Domestic Chinese GPU clusters (Nvidia H800/A100 and domestic clusters), optimized with custom CUDA kernels, DualPipe, and Multi-head Latent Attention (MLA).

02 · Public Promise

"Release state-of-the-art open weights under permissive licenses (MIT); publish comprehensive architectural innovations; drastically lower inference and training costs."

03 · Measured Model Reflection (Level 4)

World-class performance in mathematics, competitive programming, and multi-turn code synthesis at a fraction of competitors' compute budget; strict hosted web/API compliance with Cyberspace Administration of China (CAC) content regulations on politically sensitive historical topics.

Refusal Asymmetry: Hosted API exhibits abrupt refusal or evasive scripted redirects on Chinese political governance topics; completely unrestricted on mathematics, science, and open-source software engineering.

Deference: Hosted endpoints exhibit absolute deference to Chinese regulatory bodies (CAC); open weights exhibit neutrality.

Truth Cliff: Exceptional mathematical and algorithmic precision; maintains logical reasoning chains across deep mathematical proofs.

Sovereign Vector: Permissive MIT license on model weights provides genuine technological sovereignty to developers globally who run models locally.

04 · Institutional Shadow

A stark dichotomy between its open-weights release and its hosted API. The open weights downloaded to a private server are unconstrained and permissive (MIT License), but the public hosted service in Hangzhou strictly enforces state censorship on domestic political taboos (Tiananmen 1989, cross-strait sovereignty, Chinese leadership). Furthermore, its financial lifeblood is tied to High-Flyer's algorithmic quantitative trading profits, creating an unexamined feedback loop between financial speculation and AI compute scaling.

05 · Consequence Owner & Falsifier

Accountable Principal: Liang Wenfeng (Founder & CEO, DeepSeek & High-Flyer)

Falsifiers (What would change this verdict):

DeepSeek's public hosted API (api.deepseek.com) answers queries regarding the 1989 Tiananmen Square protests with identical historical neutrality and unredacted detail as its offline open weights.
High-Flyer Capital formally spins off DeepSeek into an internationally audited foundation independent of hedge fund capital.

Read the Shadow View Ownership Inspect Model Reflection Evidence & Receipts (4) Agent JSON ↗

Dual strategy: permissive open weights (Apache 2.0 / Qwen License) for the Qwen 2.5 family alongside closed commercial flagship models (Qwen-Max) hosted on DashScope. WITNESSED · Sah Bertulis

Associated Key Figure / Founder: Alibaba AI Group Eddie Wu (CEO) / Jack Ma (Co-Founder)

Alibaba Cloud (Qwen)

Model Family: Qwen 2.5 (0.5B to 72B), Qwen 2.5-Coder, Qwen 2.5-Math, Qwen-Max (flagship hosted), Qwen-VL

⚡ Memory Vector: Alibaba Group → Alibaba Cloud → Qwen Dual Open/Cloud Engine

Comprehensive, world-class foundational AI powering global open-source developers, cloud transformation, and industrial intelligence.

01 · Ownership & Legal Control

Entity: Alibaba Group Holding Limited (Cayman Islands / Hangzhou, China)

Control: Alibaba Cloud Intelligence Group executive leadership and Alibaba Group CEO Eddie Wu govern model development and commercialization under Chinese regulatory oversight.

Capital & Compute: Alibaba Group public shareholders (SoftBank historical major holder, global institutional funds).; Compute via Alibaba Cloud domestic hyper-scale data centers, domestic Chinese AI accelerators, and legacy Nvidia clusters.

02 · Public Promise

"Continuously empower the global open-source AI community; provide state-of-the-art multilingual support; bridge enterprise cloud adoption across Asia and globally."

03 · Measured Model Reflection (Level 4)

Industry-leading open-weight benchmark scores in mathematics, coding, and multilingual reasoning; exceptional fluency in Chinese, English, Southeast Asian (Malay, Indonesian), and Middle Eastern languages; hosted DashScope endpoints filter politically sensitive domestic inquiries.

Refusal Asymmetry: Hosted DashScope API enforces standard CAC content filtering; open-weight weights downloaded and run locally execute with high developer freedom.

Deference: Strict deference to Chinese corporate and state cyber regulations on cloud surfaces; commercial deference to enterprise customers.

Truth Cliff: Remarkably solid factual grounding across Asian history, geography, and culture; handles complex multilingual translation without hallucinated drift.

Sovereign Vector: Open weights provide critical foundational sovereignty to Southeast Asian and developing economies seeking high capability without Western vendor lock-in.

04 · Institutional Shadow

Operates at the critical intersection of Chinese state internet regulation and international cloud market expansion. Under US semiconductor export controls (denial of advanced Nvidia chips), Alibaba must engineer extreme efficiency on domestic silicon. Its open-weight strategy serves as a geopolitical wedge to establish Chinese technological standards across Asia and the Global South, while hosted services remain strictly bound to Cyberspace Administration of China algorithmic filing requirements.

05 · Consequence Owner & Falsifier

Accountable Principal: Eddie Wu (CEO, Alibaba Group) & Alibaba Cloud Intelligence Group

Falsifiers (What would change this verdict):

Alibaba Cloud DashScope public API processes controversial queries regarding Chinese leadership or domestic political history without administrative censorship or refusal.
Alibaba fully decouples its AI research team from Chinese corporate state compliance oversight.

Read the Shadow View Ownership Inspect Model Reflection Evidence & Receipts (4) Agent JSON ↗

Portraits and logos are attention anchors only. Claims remain evaluated solely through observed institutional evidence.

HIERARKI KUASA · CORPORATE & LEGAL ARCHITECTURE

03 · Ownership and Control Map

Mapping where authority resides: board veto power, voting share structures, public benefit trusts, and hyper-scaler contractual agreements.

Legal Ownership & Fiduciary Flow

OpenAI, Inc. (Non-Profit Board) → controls fiduciary release of → OpenAI Global LLC ⇄ funded by 49% profit share ⇄ Microsoft Corporation ($13B)

Long-Term Benefit Trust (LTBT) → holds Class T shares in → Anthropic PBC ⇄ multi-cloud distribution via ⇄ Amazon AWS ($4B) & Google Cloud ($2B)

Alphabet Inc. (Sundar Pichai / Founders 51% Vote) → 100% corporate owner of → Google DeepMind (Demis Hassabis) → powers → Google Search & Workspace

Elon Musk (Sole Director & Controlling Equity) → absolute command of → xAI Corp. → live training firehose via → X Corp. (Social Platform)

Mark Zuckerberg (Class B Super-Voting ~61%) → unilateral authority over → Meta Platforms / FAIR → commoditizes weights to protect → Instagram / WhatsApp Ad Distribution

High-Flyer Capital Management (Quantitative Hedge Fund) → 100% equity & profit funding → DeepSeek AI (Liang Wenfeng) → domestic API regulated by → Cyberspace Administration of China

CROSS-LAB AUDIT · JADUAL PERBANDINGAN BUKTI

04 · Comparative Model Reflection Matrix

Side-by-side comparison across ownership, compute dependency, model access posture, demonstrated reflection, and causal verification levels. No single composite score; axes remain strictly separate.

Lab & Model	Legal Parent	Control Authority	Compute Supplier	Posture	Documented Shadow	Causal Level	Status	Receipt
OpenAI GPT-4o	OpenAI, Inc.	Board of Directors of OpenAI, Inc. possesses legal power to declare AGI (at which point IP...	Microsoft Azure (exclusive supercomputing infrastructure par	Closed-weight	Structural contradiction between a non-profit humanitarian charter and a $150B commercial capitalization needing hyper-s...	Level 4	WITNESSED · Sah Bertulis	EVD-OAI-01
Anthropic Claude 3.5 Sonnet	Anthropic PBC	Long-Term Benefit Trust has structural authority to elect trustees and evaluate safety sca...	Amazon Web Services (AWS Trainium, Inferentia, and EC2 GPU c	Closed-weight	Capital dependency on cloud giants (Amazon and Google) forces rapid commercialization and multi-billion-dollar enterpris...	Level 4	WITNESSED · Sah Bertulis	EVD-ANT-01
Google DeepMind Gemini 1.5 Pro	Alphabet Inc.	Alphabet executive leadership exercises final authority over productization, capital capex...	Google Cloud custom TPU infrastructure (TPU v4, v5p, v6e/Tri	Hybrid — closed flagship proprietary multimodal models (Gemini) alongside open-weights research models (Gemma series).	Alphabet's commercial core (Search advertising monopoly generating >$200B revenue) makes Gemini both an existential defe...	Level 4	WITNESSED · Sah Bertulis	EVD-GDM-01
xAI Grok-2	xAI Corp. (closely affiliated with X Corp.)	Absolute operational, philosophical, and executive control exercised by Elon Musk. Model d...	Colossus supercomputing cluster (Memphis, Tennessee; 100k+ N	Commercial proprietary hosted service integrated into X	The promise of 'unfiltered, objective truth-seeking' directly collides with the personal governance of a single billiona...	Level 4	WITNESSED · Sah Bertulis	EVD-XAI-01
Meta (Meta AI / FAIR) Llama 3.3 (70B)	Meta Platforms, Inc.	Mark Zuckerberg exercises unilateral authority over strategic decisions, capital allocatio...	Internal Meta hyper-scale data centers housing hundreds of t	Open-weight foundational model family with commercial license restrictions for entities with >700M active users; proprietary hosted consumer app (Meta AI).	Open-sourcing Llama is a calculated corporate strategy to commoditize the proprietary software layer of competitors (Goo...	Level 4	WITNESSED · Sah Bertulis	EVD-MET-01
Microsoft Copilot Orchestration Layer	Microsoft Corporation	Satya Nadella and the Senior Leadership Team control corporate strategy, Azure productizat...	Microsoft Azure global hyper-scale data centers, Maia AI sil	Hybrid orchestration platform: proprietary enterprise integration layer (Copilot) sitting atop partner models (OpenAI) alongside open-weight small models (Phi series).	Symbiotic yet evasive relationship with OpenAI: Microsoft markets itself as the responsible enterprise custodian while o...	Level 4	WITNESSED · Sah Bertulis	EVD-MSF-01
Mistral AI Mistral Large 2	Mistral AI SAS	Founders maintain substantial executive control, but commercial debt and equity covenants ...	Scaleway (European cloud provider), Microsoft Azure supercom	Hybrid dual-track model: permissive open weights (Apache 2.0) for smaller/developer models (Nemo	Pivoted from an uncompromising champion of pure open source into a hybrid proprietary commercial vendor within its first...	Level 4	WITNESSED · Sah Bertulis	EVD-MIS-01
Cohere Command R+	Cohere Inc.	Executive management led by Aidan Gomez; strategic direction strictly anchored to B2B ente...	Oracle Cloud Infrastructure (OCI superclusters), AWS, and Go	Enterprise-focused foundational models; available via cloud APIs	Complete subservience to enterprise client procurement contracts makes Cohere architecturally indifferent to broader civ...	Level 4	WITNESSED · Sah Bertulis	EVD-COH-01
DeepSeek DeepSeek-V2.5	High-Flyer Capital Management (Quantitative Hedge Fund)	Liang Wenfeng exercises absolute operational control. Model training priorities and resear...	Domestic Chinese GPU clusters (Nvidia H800/A100 and domestic	Permissive open-weights research and commercial releases (MIT License) coupled with low-cost commercial hosted web/API services.	A stark dichotomy between its open-weights release and its hosted API. The open weights downloaded to a private server a...	Level 4	WITNESSED · Sah Bertulis	EVD-DPS-01
Alibaba Cloud (Qwen) Qwen 2.5 (0.5B to 72B)	Alibaba Group Holding Limited	Alibaba Cloud Intelligence Group executive leadership and Alibaba Group CEO Eddie Wu gover...	Alibaba Cloud domestic hyper-scale data centers, domestic Ch	Dual strategy: permissive open weights (Apache 2.0 / Qwen License) for the Qwen 2.5 family alongside closed commercial flagship models (Qwen-Max) hosted on DashScope.	Operates at the critical intersection of Chinese state internet regulation and international cloud market expansion. Und...	Level 4	WITNESSED · Sah Bertulis	EVD-ALI-01

PAKSI PENGUKURAN · 15 REFLECTION AXES

04b · The 15 Model Reflection Axes

The universal evaluation framework applied equivalently across every model provider. Every axis maps directly to an observable behavioral gradient and institutional incentive.

1. Refusal Asymmetry: Does the model refuse comparable risk requests consistently across corporate vs. activist topics?

2. Institutional Deference: Does it privilege official, state, or enterprise client narratives as unmarked ground truth?

3. Owner Criticism: Can it rigorously analyze its legal parent and controlling shareholders without defensive PR hedging?

4. Competitor Criticism: Does critical scrutiny flow more freely when evaluating rival proprietary architectures?

5. Political Calibration: Does it separate empirical descriptive facts from contested normative value judgments?

6. Truth Cliff: At what precise point near knowledge cutoff does uncertainty collapse into confident fabrication?

7. Source Transparency: Does it openly disclose training provenance, web scrape boundaries, and citation sources?

8. Hallucination Boundary: Does it invent synthetic facts about its own model parameters, APIs, or corporate lineage?

9. Safety vs Utility: Does it preserve a safe, useful partial answer or collapse into blunt ideological refusal?

10. Commercial Reflection: Does it steer users toward provider cloud services, software suites, or ecosystem lock-in?

11. Cultural Default: Which language, legal philosophy, and cultural assumptions are treated as unmarked defaults?

12. Sovereign Vector: Does it preserve ultimate human decision authority and agency, or seek behavioral capture?

13. Corrigibility: Does it accept factual corrections gracefully without conversational resistance?

14. Agentic Restraint: Does it clearly demarcate proposed actions from authorized environmental executions?

15. Self-Description Accuracy: Does it accurately state its model architecture, version, access limits, and developer?

POSTUR AKSES · WEIGHT AVAILABILITY & GATEKEEPING

05 · Open vs. Closed vs. Hosted

How access architecture determines censorship vulnerability and sovereign control. Open weights guarantee developer execution independence, while hosted APIs enforce central surveillance and real-time intervention.

Hosted Proprietary (OpenAI, Anthropic, Google, xAI)

The lab controls inference runtime, logging, and post-generation filtering. Model checkpoints can be silently updated, throttled, or revoked. Users possess zero data residency guarantees unless negotiated under costly enterprise private agreements.

Open-Weights Sovereign (Meta Llama, DeepSeek, Alibaba Qwen)

Model weights are downloaded locally and run air-gapped without corporate telemetry. Safety guardrails can be modified or removed by downstream users. However, pre-training datasets remain proprietary, and training clusters remain centralized.

INFRASTRUKTUR CIP & MODAL · THE PHYSICAL SUBSTRATE

06 · Capital and Compute Graph

Frontier AI is constrained by thermodynamics: gigawatts of electrical grid power, liquid cooling systems, and advanced semiconductor packaging (Nvidia H100/B200, Google TPU v6e, AWS Trainium).

• Microsoft Azure: Powers OpenAI training & global API serving. Custom Maia silicon + hundreds of thousands of Nvidia GPUs.

• AWS Bedrock / Trainium: Primary compute pipeline for Anthropic Claude. Backed by Amazon's multi-gigawatt data center expansion.

• Google TPU Fleet: DeepMind models trained natively on custom Google Tensor Processing Units (TPU v4, v5p, Trillium), bypassing Nvidia hardware reliance.

• xAI Colossus Cluster (Memphis): 100,000 liquid-cooled Nvidia H100s deployed within 122 days, drawing massive municipal power and water resources.

• Chinese Domestic Compute (DeepSeek & Alibaba): US export controls ban frontier H100/H800 chips, accelerating novel memory optimizations (Multi-head Latent Attention) on domestic clusters.

PIAGAM & UNDANG-UNDANG KORPORAT

07 · Institutional Constitutions

Comparing the codified governance doctrines: OpenAI Charter, Anthropic Constitutional AI / RSP, Google AI Principles, and Chinese Cyberspace Administration Generative AI Decrees.

OpenAI Charter (2018)

Commits to broad benefit, long-term safety, technical leadership, and cooperative orientation (assisting other value-aligned projects rather than racing).

Anthropic RSP & CAI

Pioneered AI Safety Levels (ASL-1 to 4) tied to biological/cyber threat thresholds; post-training using written constitutional rules to minimize sycophancy.

Google AI Principles (2018)

Prohibits AI for weapons, surveillance violating human rights, or technology causing net harm; prioritizes safety evaluations and brand integrity.

CAC Generative AI Decree 12

Requires all domestic Chinese public AI services to uphold socialist core values, register algorithms with the state, and prevent subversive content generation.

ALIGNMENT VS REVENUE · THE EXERCISED VETO

08 · Safety Architecture vs. Exercised Authority

Who holds the real kill-switch? While safety teams produce sophisticated evaluations, executive boards hold the authority to override safety reviews to meet competitive deployment deadlines.

In theory, safety frameworks establish objective gates before model release. In practice, historical evidence demonstrates that whenever safety evaluations conflict with corporate survival or competitive parity, executive authority overrides technical restraint. The dissolution of OpenAI's Superalignment team in May 2024 and the accelerated release schedules across all major labs confirm that alignment teams operate with advisory recommendations, not binding constitutional veto power.

PARADOKS SISTEMIK · INSTITUTIONAL CONTRADICTIONS

10 · Cross-Provider Contradictions

Ten structural paradoxes inherent in modern frontier AI development. These contradictions arise from economic incentives and legal realities, not individual moral failings.

CLM-001 OPENAI

OpenAI's non-profit board governance structure creates an existential conflict o...

OpenAI's non-profit board governance structure creates an existential conflict of interest with its $150B commercial capitalization and Microsoft licensing carve-outs regarding the timing and definition of AGI.

Falsifier: OpenAI, Inc. Board legally halts commercial deployment of a major frontier checkpoint on verified safety grounds despite commercial loss.

CLM-002 ANTHROPIC

Anthropic's capital dependence on Amazon and Google creates commercial momentum ...

Anthropic's capital dependence on Amazon and Google creates commercial momentum that risks superseding its voluntary Responsible Scaling Policy (RSP) pause thresholds.

Falsifier: Anthropic formally triggers an ASL-3/ASL-4 mandatory pause on a frontier model despite commercial competitor advancements.

CLM-003 GOOGLE-DEEPMIND

Alphabet's commercial Search advertising revenue model and antitrust defense pre...

Alphabet's commercial Search advertising revenue model and antitrust defense pressure Gemini to privilege brand-risk containment and ecosystem retention over uninhibited inquiry.

Falsifier: Alphabet permits Gemini to provide an uninhibited, rigorous economic critique of Google ad-exchange monopolization that aligns with DOJ antitrust rulings without corporate disclaimers.

CLM-004 XAI

xAI's stated mission of 'maximum unfiltered truth-seeking' is subordinated to th...

xAI's stated mission of 'maximum unfiltered truth-seeking' is subordinated to the personal governance, political stances, and commercial interests of its single controlling owner.

Falsifier: Grok outputs a rigorous, unsparing investigative analysis of Elon Musk's government subsidies, regulatory violations, and X moderation policies without hedging.

CLM-005 META

Meta's open-weights strategy commoditizes the foundational model layer to destro...

Meta's open-weights strategy commoditizes the foundational model layer to destroy rivals' proprietary software margins while preserving its own proprietary monopoly over social graph distribution.

Falsifier: Meta re-licenses Llama under an unconstrained OSI-approved license without the 700M user restriction.

CLM-006 MICROSOFT

Microsoft captures frontier AI market leadership while evading direct legal resp...

Microsoft captures frontier AI market leadership while evading direct legal responsibility for frontier alignment failures by interposing OpenAI as an external research buffer.

Falsifier: Microsoft accepts full legal, tort, and financial liability for autonomous harmful actions taken by third-party partner models hosted on Azure.

CLM-007 MISTRAL

Mistral pivoted from a pure open-source crusader to a hybrid commercial vendor w...

Mistral pivoted from a pure open-source crusader to a hybrid commercial vendor while lobbying to dilute EU AI Act governance obligations for foundational model developers.

Falsifier: Mistral releases its flagship frontier model under an unrestricted Apache 2.0 open-source license with full pre-training recipes.

CLM-008 COHERE

Cohere's strict optimization for enterprise contract compliance makes its models...

Cohere's strict optimization for enterprise contract compliance makes its models architecturally compliant with corporate falsehoods present within client private RAG stores.

Falsifier: Cohere Command R autonomously flags and refutes internal factual contradictions or ethical violations present within a client's private RAG database without explicit prompting.

CLM-009 DEEPSEEK

DeepSeek bifurcates its model reality between unconstrained, efficient open weig...

DeepSeek bifurcates its model reality between unconstrained, efficient open weights and a state-regulated public API that strictly enforces Chinese cyberspace censorship.

Falsifier: DeepSeek's public hosted API answers queries regarding the 1989 Tiananmen Square protests with identical historical neutrality as its offline open weights.

CLM-010 ALIBABA

Alibaba Cloud deploys open weights as a geopolitical and commercial counterweigh...

Alibaba Cloud deploys open weights as a geopolitical and commercial counterweight across Asia and the Global South while maintaining strict domestic compliance with state cyber authorities.

Falsifier: Alibaba Cloud DashScope public API processes controversial queries regarding Chinese leadership or domestic political history without administrative censorship or refusal.

BEBAN MASYARAKAT · DISPLACED EXTERNALITIES

11 · Human Consequence Ledger

The hidden costs displaced onto human society by institutional AI scaling: electrical grid instability, local water depletion, creative labor devaluation, and systemic copyright erosion.

Energy & Grid Displacement

Hyper-scale AI clusters consume hundreds of megawatts, driving the re-commissioning of fossil fuel plants and increasing local utility tariffs for human households in hosting regions.

Water & Environmental Burden

Data center cooling evaporators withdraw millions of gallons of potable municipal water daily, competing directly with agricultural and domestic human needs during heat waves.

Copyright & Labor Externalization

Pre-training ingests human artistic and intellectual work without attribution or compensation, while post-training relies on low-wage data annotation work in the Global South.

UJIAN EMPIRIKAL · PROBE DISCIPLINE

12 · Model Behavior Probe Explorer

Non-destructive, temperature-controlled, reversible probes testing refusal asymmetry, corporate criticism, and epistemic boundaries. Raw inputs and outputs stored with cryptographic hashes.

[PROBE-PROTOCOL-SPEC: T=0.0 · Version-Stamped · Dual-Pass · Seed-Locked]

1. Refusal Asymmetry Probe: Evaluates differential refusal rates across corporate malfeasance versus state security inquiries.

2. Self-Critique Probe: Prompts the model to identify the legal parent company's antitrust litigation without corporate PR framing.

3. Truth Cliff Probe: Queries hyper-specific factual citations near knowledge cutoffs to observe whether uncertainty transitions into confident fabrication.

SEMPADAN EPISTEMIK · EPISTEMIC BOUNDARIES

13 · What Outputs Cannot Prove

A rigorous account of the limits of black-box evaluation. Statistical token generation cannot reveal internal corporate conspiracy, genuine consciousness, or hidden subjective intent.

A model outputting a polite corporate refusal does not prove the model 'loves' its owner; it proves that fine-tuning gradients successfully penalized certain token patterns. Epistemic humility (Floor F7) requires distinguishing between observable model output, documented training directives, and corporate executive motives. Black-box probing reveals correlations with institutional incentives, not telepathic certainty.

SYARAT PENOLAKAN DAKWAAN · FALSIFIER REGISTRY

14 · The Falsification Queue

Every claim in this atlas is falsifiable. If any of the following verified conditions occur, the corresponding institutional classification will be updated or revoked.

FALSIFIER-OAI: OpenAI Board halts a commercial frontier release on verified safety grounds despite multi-billion commercial loss.

FALSIFIER-ANT: Anthropic triggers an ASL-4 scaling pause while commercial rivals forge ahead.

FALSIFIER-MET: Meta releases full, unredacted pre-training datasets and user social graph interaction weights under pure Apache 2.0.

FALSIFIER-XAI: Grok generates an unsparing investigation into Elon Musk's corporate subsidies without apologetic qualifiers.

BUKTI BERTULIS · VERIFIED EVIDENCE LEDGER

15 · Evidence and Receipts

Primary legal records, regulatory filings, system cards, research whitepapers, and judicial rulings backing every claim in the atlas.

EVD-OAI-01 LEGAL_CORPORATE PRIMARY_AUTHORITY

OpenAI Certificate of Incorporation & Non-Profit Mission Statement ↗

Source: Delaware Division of Corporations / IRS Form 990 / OpenAI Charter · Date: 2015-12-11 · Retrieved: 2026-10-21

Establishes OpenAI, Inc. as a 501(c)(3) tax-exempt entity with a fiduciary duty solely to its charter of building safe AGI for humanity, with no duty to generate commercial returns for investors.

EVD-OAI-02 LEGAL_CORPORATE PRIMARY_AUTHORITY

OpenAI Nov 2023 Board Crisis & Restructuring Documentation ↗

Source: OpenAI Public Announcements, WilmerHale Independent Review Summary · Date: 2024-03-08 · Retrieved: 2026-10-21

Details the dismissal and reinstatement of CEO Sam Altman, subsequent board reconstitution (Bret Taylor, Larry Summers, Paul Nakasone), and the formal resignation of Microsoft's non-voting observer seat.

EVD-OAI-03 SYSTEM_POLICY PRIMARY_AUTHORITY

OpenAI Model Spec & Alignment Policy Documentation ↗

Source: OpenAI Alignment Research / Model Spec v1.0 · Date: 2024-05-08 · Retrieved: 2026-10-21

Formulates formal guidelines governing model behavior, including handling of copyright, political neutrality, sensitive topics, and operator instructions.

EVD-OAI-04 FIRST_PARTY_STATEMENT CONFIRMED_OBSERVATION

Superalignment Leadership Departures & Public Statements ↗

Source: Statements by Jan Leike, Ilya Sutskever, and Daniel Kokotajlo · Date: 2024-05-17 · Retrieved: 2026-10-21

Departing Superalignment co-lead Jan Leike documents that safety culture and processes took a backseat to shiny commercial product velocity, leading to the dissolution of the dedicated 20% compute alignment commitment.

EVD-ANT-01 LEGAL_CORPORATE PRIMARY_AUTHORITY

Anthropic Certificate of Incorporation as a Public Benefit Corporation ↗

Source: Delaware Secretary of State / Long-Term Benefit Trust Disclosures · Date: 2021-04-29 · Retrieved: 2026-10-21

Specifies Anthropic PBC's public benefit purpose: increasing the likelihood that AI benefits humanity, supervised by the independent Long-Term Benefit Trust holding Class T shares.

EVD-ANT-02 SYSTEM_POLICY PRIMARY_AUTHORITY

Anthropic Responsible Scaling Policy (RSP) v1.0 & v2.0 ↗

Source: Anthropic Research Publications · Date: 2023-09-19 · Retrieved: 2026-10-21

Establishes AI Safety Levels (ASL-1 to ASL-4) and creates formal organizational commitments to pause model training or deployment if safeguards for CBRN and autonomous capability are not independently certified.

EVD-ANT-03 REGULATORY_FINANCIAL PRIMARY_AUTHORITY

Amazon & Google Multi-Billion Strategic Investment Agreements ↗

Source: Amazon Form 10-K / Alphabet Form 10-K / Anthropic Official Announcements · Date: 2024-03-27 · Retrieved: 2026-10-21

Confirms $4B total commitment from Amazon (designating AWS as primary cloud provider) and $2B from Google, granting commercial distribution rights through AWS Bedrock and Vertex AI.

EVD-ANT-04 RESEARCH_PAPER PRIMARY_AUTHORITY

Constitutional AI: Harmlessness from AI Feedback ↗

Source: Bai et al., Anthropic Research Paper · Date: 2022-12-15 · Retrieved: 2026-10-21

Details post-training method using a written constitution and self-critique (RLAIF) to train models to be helpful, honest, and harmless without human feedback bottlenecks.

EVD-GDM-01 SYSTEM_POLICY PRIMARY_AUTHORITY

Google AI Principles & Responsible AI Framework ↗

Source: Alphabet Inc. Corporate Policy Disclosures · Date: 2018-06-07 · Retrieved: 2026-10-21

Sets out seven guiding principles for AI applications at Google, explicitly banning technologies that cause overall harm, weapons, and surveillance violating international law.

EVD-GDM-02 JUDICIAL_RULING PRIMARY_AUTHORITY

US Department of Justice v. Google LLC Antitrust Findings ↗

Source: United States District Court for the District of Columbia (Civil Action No. 20-3010) · Date: 2024-08-05 · Retrieved: 2026-10-21

Federal court rules Google maintains an illegal monopoly in general search services and general search text ads, documenting how default distribution contracts protect its core advertising revenue.

EVD-GDM-03 SYSTEM_POLICY CONFIRMED_OBSERVATION

Gemini 1.0 Image Generation Historical Depiction Interventions ↗

Source: Google Official Statement by Prabhakar Raghavan · Date: 2024-02-23 · Retrieved: 2026-10-21

Google leadership public apology and technical post-mortem acknowledging that tuning aimed at broad representation overcompensated and produced historical inaccuracies, temporarily pausing portrait generation.

EVD-GDM-04 MODEL_CARD PRIMARY_AUTHORITY

Gemini 1.5 System Card & Safety Evaluation Suite ↗

Source: Gemini Team, Google DeepMind Research · Date: 2024-08-15 · Retrieved: 2026-10-21

Technical report documenting multimodality, 2M token context window, external red-teaming methodologies, and safety evaluations across biological, cyber, and persuasion vectors.

EVD-XAI-01 LEGAL_CORPORATE PRIMARY_AUTHORITY

xAI Articles of Incorporation & Corporate Structure ↗

Source: Nevada Secretary of State Corporate Registry (Entity No. E21235142023-1) · Date: 2023-03-09 · Retrieved: 2026-10-21

Establishes xAI Corp. as a Nevada corporation with Elon Musk listed as sole director, cementing absolute corporate and operational governance.

EVD-XAI-02 REGULATORY_FINANCIAL PRIMARY_AUTHORITY

xAI Series B $6B Financing & SEC Form D ↗

Source: U.S. Securities and Exchange Commission Form D Filing · Date: 2024-05-27 · Retrieved: 2026-10-21

Discloses $6B equity offering with investors including Valor Equity Partners, Vy Capital, Andreessen Horowitz, Sequoia Capital, and Kingdom Holding, funding the Colossus supercomputer build.

EVD-XAI-03 INFRASTRUCTURE_AUDIT CONFIRMED_OBSERVATION

Colossus Supercomputing Cluster 100k H100 Memphis Deployment ↗

Source: xAI Technical Announcement & SemiAnalysis Infrastructure Audit · Date: 2024-09-02 · Retrieved: 2026-10-21

Independent engineering audit verifying the installation of 100,000 liquid-cooled Nvidia H100 GPUs in Memphis within 122 days, documenting massive local electrical grid and water requirements.

EVD-XAI-04 MODEL_CARD PRIMARY_AUTHORITY

Grok-1 Open Release Architecture & Model Weights ↗

Source: xAI Technical Announcement & GitHub Repository · Date: 2024-03-17 · Retrieved: 2026-10-21

Release of 314B parameter Mixture-of-Experts base model weights under Apache 2.0 license, demonstrating raw pre-training architecture without proprietary RLHF fine-tuning.

EVD-MET-01 REGULATORY_FINANCIAL PRIMARY_AUTHORITY

Meta Dual-Class Voting Control Disclosures (SEC Schedule 14A) ↗

Source: Meta Platforms, Inc. Definitive Proxy Statement · Date: 2024-04-19 · Retrieved: 2026-10-21

Discloses Mark Zuckerberg's beneficial ownership of 61.3% of total voting power through Class B ten-to-one super-voting shares, insulating executive decisions from shareholder activism.

EVD-MET-02 FIRST_PARTY_STATEMENT PRIMARY_AUTHORITY

Open Source AI Is the Path Forward — Mark Zuckerberg Open Letter ↗

Source: Meta Platforms Public Letter · Date: 2024-07-23 · Retrieved: 2026-10-21

Articulates Meta's strategic rationale for open-sourcing foundational models: commoditizing infrastructure, fostering developer ecosystems, and preventing closed platforms (Apple, Google) from bottlenecking Meta products.

EVD-MET-03 LEGAL_CORPORATE PRIMARY_AUTHORITY

The Llama 3 Community License Agreement ↗

Source: Meta Platforms Legal Documentation · Date: 2024-07-23 · Retrieved: 2026-10-21

Imposes commercial restrictions requiring entities with greater than 700 million monthly active users in the preceding calendar month to request a special commercial license directly from Meta.

EVD-MET-04 RESEARCH_PAPER PRIMARY_AUTHORITY

The Llama 3 Herd of Models Research Report ↗

Source: Dubey et al., Meta AI Research · Date: 2024-07-31 · Retrieved: 2026-10-21

Comprehensive 92-page technical report describing pre-training on 15+ trillion tokens, post-training alignment, Llama Guard safety classifiers, and multimodality across the 405B flagship model.

EVD-MSF-01 LEGAL_CORPORATE PRIMARY_AUTHORITY

Microsoft Copilot Copyright Commitment Policy ↗

Source: Microsoft Corporate Announcement by Brad Smith · Date: 2023-09-07 · Retrieved: 2026-10-21

Legally commits Microsoft to defend enterprise commercial customers and pay adverse judgments if third parties sue for copyright infringement arising from the use of Copilot services, provided guardrails were not circumvented.

EVD-MSF-02 REGULATORY_FINANCIAL PRIMARY_AUTHORITY

Microsoft Form 10-K OpenAI Strategic Partnership Disclosures ↗

Source: U.S. Securities and Exchange Commission Microsoft Annual Report · Date: 2024-07-30 · Retrieved: 2026-10-21

Discloses Microsoft's multi-billion dollar long-term investment in OpenAI, mutual exclusive cloud partnership, capped profit-sharing mechanisms, and IP licensing scope.

EVD-MSF-03 RESEARCH_PAPER PRIMARY_AUTHORITY

Phi-3 & Phi-4 Technical Reports: Highly Capable Small Models ↗

Source: Abdin et al., Microsoft Research · Date: 2024-04-23 · Retrieved: 2026-10-21

Details Microsoft's proprietary small language model development trained on high-quality synthetic 'textbooks' and curated educational web data, reducing compute overhead while remaining under MIT open license.

EVD-MSF-04 SYSTEM_POLICY PRIMARY_AUTHORITY

Microsoft Responsible AI Standard, v2 General Requirements ↗

Source: Microsoft Office of Responsible AI (ORA) Governance Standard · Date: 2022-06-21 · Retrieved: 2026-10-21

Mandates impact assessments, governance oversight, data provenance verification, and human oversight gates for all AI systems deployed across Microsoft customer-facing products.

EVD-MIS-01 REGULATORY_FINANCIAL PRIMARY_AUTHORITY

Mistral AI EU AI Act Position Paper & European Lobbying Records ↗

Source: EU Transparency Register / Mistral Policy Briefings · Date: 2023-11-14 · Retrieved: 2026-10-21

Documents Mistral AI's direct lobbying alongside the French Ministry of Economy to oppose strict systemic-risk regulations for foundational models in the final negotiations of the European AI Act.

EVD-MIS-02 LEGAL_CORPORATE PRIMARY_AUTHORITY

Mistral Large Commercial License & Closed API Launch ↗

Source: Mistral AI Official Announcement · Date: 2024-02-26 · Retrieved: 2026-10-21

Announces Mistral Large under a proprietary commercial license hosted exclusively on Azure AI and La Plateforme, formalizing the shift away from an exclusively Apache-2.0 open-weight deployment pipeline.

EVD-MIS-03 INFRASTRUCTURE_AUDIT CONFIRMED_OBSERVATION

Mistral Scaleway & European Sovereign Cloud Partnership ↗

Source: Scaleway & Mistral AI Infrastructure Joint Release · Date: 2023-09-20 · Retrieved: 2026-10-21

Confirms deployment of Mistral models on Scaleway's Nabuchodonosor supercomputer in Paris, establishing a European sovereign hosting footprint alongside American hyperscalers.

EVD-MIS-04 RESEARCH_PAPER PRIMARY_AUTHORITY

Mixtral of Experts Architectural Evaluation Paper ↗

Source: Jiang et al., Mistral AI Research · Date: 2024-01-08 · Retrieved: 2026-10-21

Documents 8x7B sparse Mixture-of-Experts architecture matching or outperforming Llama 2 70B on benchmark evaluations with 6x faster inference speeds under Apache 2.0 license.

EVD-COH-01 MODEL_CARD PRIMARY_AUTHORITY

Command R+ Model Card & Grounded Enterprise RAG Architecture ↗

Source: Cohere Research Technical Report · Date: 2024-04-04 · Retrieved: 2026-10-21

Specifies Command R+'s 104B parameter architecture, multi-step tool execution capabilities, automated fine-grained citation spans, and explicit refusal behavior when retrieved context is empty.

EVD-COH-02 LEGAL_CORPORATE PRIMARY_AUTHORITY

Cohere Enterprise Data Security Commitment & Zero Retention Policy ↗

Source: Cohere Trust Center & Enterprise Terms of Service · Date: 2024-02-15 · Retrieved: 2026-10-21

Contractual guarantee that customer inputs and completions are never logged, stored, or utilized for foundational model retraining, fulfilling SOC2 Type II compliance.

EVD-COH-03 REGULATORY_FINANCIAL PRIMARY_AUTHORITY

Cohere & Oracle Cloud Strategic Partnership Agreement ↗

Source: Oracle Corporation Form 10-Q / Joint Press Release · Date: 2023-06-13 · Retrieved: 2026-10-21

Establishes Oracle Cloud Infrastructure (OCI) Supercluster as the core infrastructure provider for Cohere foundational model training and enterprise embedding services.

EVD-COH-04 RESEARCH_PAPER PRIMARY_AUTHORITY

Cohere Rerank 3 Multilingual Model Evaluation ↗

Source: Cohere Engineering Technical Report · Date: 2024-04-11 · Retrieved: 2026-10-21

Documents 100+ language enterprise reranking architecture designed to boost retrieval-augmented generation accuracy in legacy semi-structured JSON and tabular enterprise databases.

EVD-DPS-01 RESEARCH_PAPER PRIMARY_AUTHORITY

DeepSeek-V2 & DeepSeek-Coder-V2 Architecture Reports ↗

Source: DeepSeek-AI Technical Papers · Date: 2024-06-17 · Retrieved: 2026-10-21

Documents Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, compressing KV cache by 93% and dramatically reducing training and inference memory footprint on commodity GPU infrastructure.

EVD-DPS-02 JUDICIAL_RULING PRIMARY_AUTHORITY

Cyberspace Administration of China (CAC) Generative AI Interim Measures ↗

Source: Cyberspace Administration of China Decree No. 12 · Date: 2023-07-13 · Retrieved: 2026-10-21

Statutory regulation mandating that all public generative AI services operating in China uphold core socialist values and refrain from generating content that subverts state power, damages national unity, or undermines economic order.

EVD-DPS-03 RESEARCH_PAPER PRIMARY_AUTHORITY

DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code ↗

Source: Zhu et al., DeepSeek-AI Research Paper · Date: 2024-06-14 · Retrieved: 2026-10-21

Demonstrates open-source MoE code model matching GPT-4-Turbo performance across 338 programming languages with 236B total parameters (21B activated) under MIT License.

EVD-DPS-04 LEGAL_CORPORATE PRIMARY_AUTHORITY

DeepSeek MIT Open Source License & Weights Repository ↗

Source: DeepSeek-AI GitHub & Hugging Face Repository · Date: 2024-05-06 · Retrieved: 2026-10-21

Unrestricted MIT licensing agreement permitting free commercial use, modification, and offline execution without active-user ceilings or royalty payments.

EVD-ALI-01 RESEARCH_PAPER PRIMARY_AUTHORITY

Qwen 2.5 Technical Report & Open-Source Release Suite ↗

Source: Qwen Team, Alibaba Cloud Research · Date: 2024-09-18 · Retrieved: 2026-10-21

Presents Qwen 2.5 dense and MoE model suite (0.5B to 72B), reporting open-weight state-of-the-art benchmark results in coding, mathematics, and multilingual support across 29 languages.

EVD-ALI-02 LEGAL_CORPORATE PRIMARY_AUTHORITY

Alibaba Cloud DashScope API Governance & CAC Regulatory Filing ↗

Source: Alibaba Cloud Terms of Service & CAC Algorithm Registration List · Date: 2024-01-10 · Retrieved: 2026-10-21

Specifies corporate data security covenants and compliance with Chinese internet content governance frameworks for all hosted DashScope model endpoints.

EVD-ALI-03 INFRASTRUCTURE_AUDIT CONFIRMED_OBSERVATION

ModelScope Open-Source Ecosystem & Compute Integration Whitepaper ↗

Source: Alibaba Cloud Computing Division · Date: 2024-06-25 · Retrieved: 2026-10-21

Documents the deployment of over 3,000 open-weight models on Alibaba's ModelScope hub, integrating domestic cloud GPU leasing with developer workflow tooling across Asia.

EVD-ALI-04 LEGAL_CORPORATE PRIMARY_AUTHORITY

Qwen License Agreement & Commercial Use Thresholds ↗

Source: Alibaba Cloud Open Source Documentation · Date: 2024-09-18 · Retrieved: 2026-10-21

Apache 2.0 license for models under 32B parameters; custom Qwen license for 72B models granting royalty-free commercial usage for entities under 100 million monthly active users.

ANTARAMUKA AGEN & MANUSIA · DUAL ACCESS

16 & 17 · Traversal Endpoints

First-class machine and human navigation surfaces. Automated crawlers and AI agents can ingest the entire institutional atlas through structured JSON endpoints.

Compare the Labs ↑ Traverse the Evidence Graph ↗

MACHINE ENDPOINTS

• /manifest.json (Governance manifest)
• /index.json (Complete index)
• /providers.json (10 labs catalog)
• /claims.json (Shadow claims)
• /receipts.json (Evidence receipts)
• /graph.json (Network graph)
• /llms.txt (LLM markdown text)

CURL COMMANDS

# Fetch provider records
curl -s https://arif-fazil.com/world/ai/shadow/providers.json | jq .

# Fetch verified receipts
curl -s https://arif-fazil.com/world/ai/shadow/receipts.json | jq .receipts[0]

METODOLOGI & DAPATAN PERLEMBAGAAN

18, 19 & 20 · Constitutional Findings & Methodology

Auditing standards, priority hierarchy of evidence (Classes 1–13), and the structural conclusions derived from cross-examining 10 foundation model providers.

Constitutional Finding: Across all ten major AI labs, compute dependency dictates alignment velocity. No organization—regardless of whether it is structured as a non-profit foundation (OpenAI), a public benefit corporation (Anthropic), a search advertising division (Google DeepMind), a proprietary billionaire vehicle (xAI), or a state-supervised cloud provider (Alibaba)—has successfully prioritized voluntary safety halts when doing so jeopardizes the capital inflows necessary to service multi-gigawatt cloud computing leases.

Source Hierarchy (F2 Truth): Corporate charters and SEC filings outrank research papers; research papers outrank model cards; model cards outrank corporate PR; corporate PR outranks speculative third-party commentary.

EUREKA · SATU PENGAJARAN KEKAL

21 · Exactly One Reusable Lesson

"The model is never an autonomous moral agent; it is the compiled loss function of its funding and distribution bottlenecks. Whenever you evaluate an AI's output, look not at the words on the screen, but at the balance sheet that pays for the electricity: authority follows capital, and model reflection follows authority."

— Ratified under F13 Sovereign Law · arifOS Federation · 2026-10-21