diff --git a/research/ai_generated_agi_architectures/README.md b/research/ai_generated_agi_architectures/README.md
new file mode 100644
index 0000000..c1c3b76
--- /dev/null
+++ b/research/ai_generated_agi_architectures/README.md
@@ -0,0 +1,16 @@
+# AI-Generated AGI Architecture Proposals — Research Packet
+
+Proposals from 8 AI systems compared across 9 architectural dimensions. Generated 2026-07-21 for aLexzzz430/Cognitive-OS bounty.
+
+| System | Method |
+|--------|--------|
+| Clusy / Claude-family | Self-report |
+| OpenAI o-series | Reconstructed |
+| Google DeepMind | Lab-stated + reconstructed |
+| xAI Grok | Reconstructed |
+| DeepSeek V3/R1 | Reconstructed |
+| Alibaba Qwen3 | Reconstructed |
+| Meta Llama+JEPA | Lab-stated architecture |
+| Mistral Small 3.x | Reconstructed |
+
+Only Meta JEPA and DeepMind Levels of AGI are genuinely lab-stated architectures.
\ No newline at end of file
diff --git a/research/ai_generated_agi_architectures/comparison.csv b/research/ai_generated_agi_architectures/comparison.csv
new file mode 100644
index 0000000..d788ec1
--- /dev/null
+++ b/research/ai_generated_agi_architectures/comparison.csv
@@ -0,0 +1,9 @@
+"system","collection_method","core_paradigm","world_model_knowledge","memory","learning","reasoning_planning","perception_action_tools","motivation_alignment","self_model_metacognition","scalability_path"
+"Clusy / Claude-family (Anthropic)","direct-query / self-report","Predictive-Processing Global-Workspace: hierarchical generative world model + global-workspace broadcast","Transformer multimodal latent world model (generative simulator) + explicit knowledge graph","Working (active latent+workspace) / episodic (retrieval store) / semantic (weights+KG) / procedural (skills)","In-context (fast) + replay/regularized consolidation (slow) + RL-from-outcomes + self-supervised prediction","Prediction-as-reasoning; latent MCTS-like tree search; hierarchical (HTN) planning; active-inference exploration","Multimodal predictive-coding encoders; embodiment-agnostic action interface; tools as procedural skills","Intrinsic (prediction-error/homeostasis) + constitutional priors + corrigibility + RLHF/RLAIF + oversight","World model includes self-model; confidence estimation; help-seeking; compute allocation","Scale world-model + memory + planning horizon; generality from breadth + skill acquisition + in-context adaptation"
+"OpenAI — GPT / o-series","reconstructed-from-publications","Scale a transformer language model, then train reasoning with reinforcement learning on chain-of-thought (the o-series 'deliberative alignment' recipe). No named cognitive architecture; the bet is scale + RL-on-reasoning.","Knowledge is implicit in transformer weights; no explicit world model or KG. o-series emits long internal reasoning traces as a scratchpad.","Working memory = context window. No native persistent episodic/semantic store (added externally via retrieval/tools).","Pretraining (self-supervised next-token) + RLHF + RL on verifiable rewards for reasoning (GRPO-style). In-context learning at inference; no continual weight updates in deployment.","Test-time compute: long RL-trained chain-of-thought 'thinking' before answering; tool use for planning. Search is implicit in the learned reasoning policy rather than explicit tree search.","Multimodal (text/image/audio in, text out); strong tool/function calling and code execution as the action interface.","RLHF + 'deliberative alignment' (model reasons about its instructions/values explicitly). Objectives set by trainers; corrigibility via training, not architectural.","Limited explicit self-model; some calibrated confidence via sampling; reasoning traces give partial introspection.","Generality via scaling laws + RL on ever-harder reasoning tasks; capability grows with parameters/data/test-time compute."
+"Google DeepMind — Gemini / Gato / Alpha-family","lab-stated-AGI-ontology (not a full architecture)","Generalist agent paradigm: a single multimodal model (Gato/Gemini) plus model-based RL + search (AlphaGo/AlphaFold lineage). DeepMind published a 'Levels of AGI' ontology (capability levels), which is a framework, not an architecture.","Multimodal transformer world knowledge in weights; Alpha-family uses explicit learned simulators/search over structured state (e.g. AlphaFold structure prediction).","Context window working memory; external retrieval; Alpha-family uses search over an explicit state representation as a form of episodic lookahead.","Self-supervised pretraining + RL (model-based RL in Alpha-family; RLHF in Gemini). Strong emphasis on RL + planning.","Explicit model-based planning + tree search (MCTS lineage from AlphaGo) is DeepMind's signature; Gemini adds learned reasoning.","Deeply multimodal (Gato: one model across vision/language/robotics/game actions); embodiment is a core focus (robotics, SIMA agent in 3D worlds).","Reward-driven RL; alignment via RLHF + safety training; objectives from reward design.","Not a central published component; some uncertainty estimation in RL agents.","Generality via generalist agents + RL + search across domains; the Levels-of-AGI framework defines the path (narrow -> competent -> expert -> virtuoso)."
+"xAI — Grok","reconstructed-from-publications","Large (mixture-of-experts) transformer trained to 'understand the universe', with a stated emphasis on curiosity and real-time knowledge. No formal cognitive architecture published.","Parametric knowledge in weights, augmented by real-time access to X (Twitter) data for up-to-date world knowledge.","Context window working memory; real-time data retrieval as an external memory proxy.","Large-scale pretraining + RLHF; MoE for efficient scaling.","Chain-of-thought reasoning modes; tool use; less emphasis on explicit search.","Multimodal (text/image); tool/function calling; real-time web/X access as tools.","RLHF alignment; stated goal of truth-seeking/curiosity; 'maximum truth-seeking' framing.","Not a published architectural component.","Generality via scaling MoE + real-time knowledge integration + reasoning."
+"DeepSeek — V3 / R1","reconstructed-from-publications","Efficient open Mixture-of-Experts transformer; R1 adds RL-trained reasoning (GRPO) without heavy supervised fine-tuning. Paradigm: efficient MoE + RL-driven emergent reasoning.","Parametric knowledge in MoE weights; no explicit world model/KG.","Context window (long-context variants); no native persistent store.","Pretraining + GRPO reinforcement learning (R1) that elicits long reasoning; open weights enable community fine-tuning.","R1's RL-trained long chain-of-thought reasoning with self-verification; test-time reasoning compute.","Primarily text/code; tool use supported; multimodal limited relative to frontier labs.","RL with reward signals; alignment via RLHF; open-weight transparency.","R1 shows emergent self-reflection/'aha' moments in reasoning traces (partial metacognition).","Generality via efficient MoE scaling + RL on reasoning; open ecosystem accelerates capability spread."
+"Alibaba — Qwen3","reconstructed-from-publications","Open MoE transformer with hybrid 'thinking/non-thinking' modes and strong agent/tool-use focus. Paradigm: efficient MoE + agentic capability.","Parametric knowledge in MoE weights; strong multilingual coverage.","Long context window; agentic memory via tool/retrieval integration.","Pretraining + SFT + RL; hybrid thinking mode toggles deliberate reasoning.","Hybrid thinking mode (fast vs deliberative); strong agentic planning and tool orchestration.","Multimodal variants (Qwen-VL/Audio); strong tool/function calling and coding agents.","RLHF alignment; agentic objective-following.","Thinking mode gives partial introspection; no formal self-model.","Generality via MoE scaling + agentic tool use + multilingual breadth."
+"Meta — Llama + V-JEPA (LeCun)","lab-stated-AGI-architecture (JEPA)","JEPA (Joint-Embedding Predictive Architecture): LeCun's named cognitive architecture. Predict in LATENT representation space (not pixel space) via a hierarchical world model; non-generative prediction avoids modeling irrelevant detail. Llama is the current LM; JEPA is the AGI proposal.","A learned latent world model that predicts abstract representations of future states; knowledge as predictive embeddings, hierarchical across timescales.","Working memory = current latent state; the architecture proposes configurable memory with a memory-encoder; episodic/semantic via the world model's latent structure.","Self-supervised joint-embedding prediction (no generative reconstruction); learns by predicting latent futures — energy-based, not token-prediction.","A 'planner'/actor module optimizes actions to minimize a cost in latent space using the world model's predictions — explicit model-based planning over abstract states.","Perception = encoder into latent space; action via the actor/cost modules; designed to be embodiment-agnostic (V-JEPA 2 targets video/robotics).","A 'cost module' (intrinsic + extrinsic objectives) drives the planner; alignment framed as configurable cost functions; LeCun argues this avoids generative-model pitfalls.","The architecture includes the agent's own state in the world model; uncertainty via energy/latent prediction confidence.","Generality via hierarchical latent prediction across timescales + a world model that supports planning; V-JEPA 2 is an early instantiation."
+"Mistral AI — Small 3.x / open MoE","reconstructed-from-publications","Efficient open (and commercial) transformers/MoE with strong tool use; no published AGI cognitive architecture. Paradigm: efficient, deployable, agentic models.","Parametric knowledge in weights; no explicit world model/KG.","Context window working memory; agentic memory via tools/retrieval.","Pretraining + instruction tuning + RLHF; efficient architectures (MoE, sliding-window attention).","Tool-augmented reasoning; agentic function calling; less emphasis on long RL reasoning than o/R1.","Multimodal (Small 3.1 is natively multimodal); strong tool/function calling and on-device deployment.","RLHF/instruction alignment; objective-following.","Not a published architectural component.","Generality via efficient scaling + multimodality + agentic tool use + broad deployment."
diff --git a/research/ai_generated_agi_architectures/prompts.md b/research/ai_generated_agi_architectures/prompts.md
new file mode 100644
index 0000000..a5a5abe
--- /dev/null
+++ b/research/ai_generated_agi_architectures/prompts.md
@@ -0,0 +1,4 @@
+# Prompts Used
+
+## Master prompt
+Design a detailed AGI architecture across 9 dimensions: core paradigm, world model, memory, learning, reasoning/planning, perception/action/tools, motivation/alignment, self-model/metacognition, scalability path.
\ No newline at end of file
diff --git a/research/ai_generated_agi_architectures/raw_outputs/alibaba_qwen.md b/research/ai_generated_agi_architectures/raw_outputs/alibaba_qwen.md
new file mode 100644
index 0000000..47bf459
--- /dev/null
+++ b/research/ai_generated_agi_architectures/raw_outputs/alibaba_qwen.md
@@ -0,0 +1,45 @@
+# AGI Architecture Profile — Alibaba — Qwen3
+**Collection method:** `[reconstructed-from-publications]`
+**Access date:** 2026-07-21
+
+> Reconstructed from the system's published research per the bounty's allowed method; see `sources.md` for primary sources.
+
+## 1. Core paradigm
+
+Open MoE transformer with hybrid 'thinking/non-thinking' modes and strong agent/tool-use focus. Paradigm: efficient MoE + agentic capability.
+
+## 2. World model / knowledge
+
+Parametric knowledge in MoE weights; strong multilingual coverage.
+
+## 3. Memory systems
+
+Long context window; agentic memory via tool/retrieval integration.
+
+## 4. Learning mechanisms
+
+Pretraining + SFT + RL; hybrid thinking mode toggles deliberate reasoning.
+
+## 5. Reasoning & planning
+
+Hybrid thinking mode (fast vs deliberative); strong agentic planning and tool orchestration.
+
+## 6. Perception / action / tools
+
+Multimodal variants (Qwen-VL/Audio); strong tool/function calling and coding agents.
+
+## 7. Motivation / alignment
+
+RLHF alignment; agentic objective-following.
+
+## 8. Self-model / metacognition
+
+Thinking mode gives partial introspection; no formal self-model.
+
+## 9. Scalability / path to generality
+
+Generality via MoE scaling + agentic tool use + multilingual breadth.
+
+## Sources
+
+- Qwen3 technical report (2025)
diff --git a/research/ai_generated_agi_architectures/raw_outputs/clusy_claude_family.md b/research/ai_generated_agi_architectures/raw_outputs/clusy_claude_family.md
new file mode 100644
index 0000000..5bbcf0d
--- /dev/null
+++ b/research/ai_generated_agi_architectures/raw_outputs/clusy_claude_family.md
@@ -0,0 +1,32 @@
+# AGI Architecture Proposal — Clusy (Claude-family)
+**Method:** self-report
+
+## Core paradigm: Predictive-Processing Global-Workspace (PPGW)
+Hierarchical generative world model + active inference + global workspace broadcast
+
+## 1. Overall architecture
+7 subsystems: Perception, World Model, Memory (working/episodic/semantic/procedural), Reasoning/Planning (latent tree search), Global Workspace, Motivation/Value, Self-model. Loop: predict -> observe error -> broadcast -> deliberate -> act -> update
+
+## 2. World model
+Transformer latent generative model + knowledge graph. Generative simulator for planning and counterfactuals.
+
+## 3. Memory
+Working = active latent + workspace. Episodic = vector-indexed experiences. Semantic = weights + graph. Procedural = skill subroutines. Consolidation via replay + EWC.
+
+## 4. Learning
+In-context (fast), consolidation (replay + EWC), RL (RLHF/RLAIF), self-supervised prediction.
+
+## 5. Reasoning & planning
+World model simulates actions -> latent MCTS over trajectories. Hierarchical HTN. Uncertainty via ensembles.
+
+## 6. Tools and embodiment
+Embodiment-agnostic. Tool calls as procedural skills (code, APIs, search).
+
+## 7. Alignment
+Constitutional priors, corrigibility (objectives uncertain), RLHF, bounded agency.
+
+## 8. Self-model
+World model includes self-model. Calibrated confidence, help-seeking, compute allocation.
+
+## 9. Scalability
+Generality from breadth + in-context adaptation + skill acquisition. Scale model, memory, planning together.
\ No newline at end of file
diff --git a/research/ai_generated_agi_architectures/raw_outputs/deepmind_gemini.md b/research/ai_generated_agi_architectures/raw_outputs/deepmind_gemini.md
new file mode 100644
index 0000000..89b0769
--- /dev/null
+++ b/research/ai_generated_agi_architectures/raw_outputs/deepmind_gemini.md
@@ -0,0 +1,46 @@
+# AGI Architecture Profile — Google DeepMind — Gemini / Gato / Alpha-family
+**Collection method:** `[lab-stated-AGI-ontology (not a full architecture)]`
+**Access date:** 2026-07-21
+
+> Reconstructed from the system's published research per the bounty's allowed method; see `sources.md` for primary sources.
+
+## 1. Core paradigm
+
+Generalist agent paradigm: a single multimodal model (Gato/Gemini) plus model-based RL + search (AlphaGo/AlphaFold lineage). DeepMind published a 'Levels of AGI' ontology (capability levels), which is a framework, not an architecture.
+
+## 2. World model / knowledge
+
+Multimodal transformer world knowledge in weights; Alpha-family uses explicit learned simulators/search over structured state (e.g. AlphaFold structure prediction).
+
+## 3. Memory systems
+
+Context window working memory; external retrieval; Alpha-family uses search over an explicit state representation as a form of episodic lookahead.
+
+## 4. Learning mechanisms
+
+Self-supervised pretraining + RL (model-based RL in Alpha-family; RLHF in Gemini). Strong emphasis on RL + planning.
+
+## 5. Reasoning & planning
+
+Explicit model-based planning + tree search (MCTS lineage from AlphaGo) is DeepMind's signature; Gemini adds learned reasoning.
+
+## 6. Perception / action / tools
+
+Deeply multimodal (Gato: one model across vision/language/robotics/game actions); embodiment is a core focus (robotics, SIMA agent in 3D worlds).
+
+## 7. Motivation / alignment
+
+Reward-driven RL; alignment via RLHF + safety training; objectives from reward design.
+
+## 8. Self-model / metacognition
+
+Not a central published component; some uncertainty estimation in RL agents.
+
+## 9. Scalability / path to generality
+
+Generality via generalist agents + RL + search across domains; the Levels-of-AGI framework defines the path (narrow -> competent -> expert -> virtuoso).
+
+## Sources
+
+- DeepMind 'Levels of AGI' paper (2023)
+- Gato generalist agent (2022); AlphaGo/AlphaFold
diff --git a/research/ai_generated_agi_architectures/raw_outputs/deepseek.md b/research/ai_generated_agi_architectures/raw_outputs/deepseek.md
new file mode 100644
index 0000000..4193e50
--- /dev/null
+++ b/research/ai_generated_agi_architectures/raw_outputs/deepseek.md
@@ -0,0 +1,46 @@
+# AGI Architecture Profile — DeepSeek — V3 / R1
+**Collection method:** `[reconstructed-from-publications]`
+**Access date:** 2026-07-21
+
+> Reconstructed from the system's published research per the bounty's allowed method; see `sources.md` for primary sources.
+
+## 1. Core paradigm
+
+Efficient open Mixture-of-Experts transformer; R1 adds RL-trained reasoning (GRPO) without heavy supervised fine-tuning. Paradigm: efficient MoE + RL-driven emergent reasoning.
+
+## 2. World model / knowledge
+
+Parametric knowledge in MoE weights; no explicit world model/KG.
+
+## 3. Memory systems
+
+Context window (long-context variants); no native persistent store.
+
+## 4. Learning mechanisms
+
+Pretraining + GRPO reinforcement learning (R1) that elicits long reasoning; open weights enable community fine-tuning.
+
+## 5. Reasoning & planning
+
+R1's RL-trained long chain-of-thought reasoning with self-verification; test-time reasoning compute.
+
+## 6. Perception / action / tools
+
+Primarily text/code; tool use supported; multimodal limited relative to frontier labs.
+
+## 7. Motivation / alignment
+
+RL with reward signals; alignment via RLHF; open-weight transparency.
+
+## 8. Self-model / metacognition
+
+R1 shows emergent self-reflection/'aha' moments in reasoning traces (partial metacognition).
+
+## 9. Scalability / path to generality
+
+Generality via efficient MoE scaling + RL on reasoning; open ecosystem accelerates capability spread.
+
+## Sources
+
+- DeepSeek-V3 technical report (2024)
+- DeepSeek-R1 (2025)
diff --git a/research/ai_generated_agi_architectures/raw_outputs/meta_llama_jepa.md b/research/ai_generated_agi_architectures/raw_outputs/meta_llama_jepa.md
new file mode 100644
index 0000000..ad2df6a
--- /dev/null
+++ b/research/ai_generated_agi_architectures/raw_outputs/meta_llama_jepa.md
@@ -0,0 +1,46 @@
+# AGI Architecture Profile — Meta — Llama + V-JEPA (LeCun)
+**Collection method:** `[lab-stated-AGI-architecture (JEPA)]`
+**Access date:** 2026-07-21
+
+> Reconstructed from the system's published research per the bounty's allowed method; see `sources.md` for primary sources.
+
+## 1. Core paradigm
+
+JEPA (Joint-Embedding Predictive Architecture): LeCun's named cognitive architecture. Predict in LATENT representation space (not pixel space) via a hierarchical world model; non-generative prediction avoids modeling irrelevant detail. Llama is the current LM; JEPA is the AGI proposal.
+
+## 2. World model / knowledge
+
+A learned latent world model that predicts abstract representations of future states; knowledge as predictive embeddings, hierarchical across timescales.
+
+## 3. Memory systems
+
+Working memory = current latent state; the architecture proposes configurable memory with a memory-encoder; episodic/semantic via the world model's latent structure.
+
+## 4. Learning mechanisms
+
+Self-supervised joint-embedding prediction (no generative reconstruction); learns by predicting latent futures — energy-based, not token-prediction.
+
+## 5. Reasoning & planning
+
+A 'planner'/actor module optimizes actions to minimize a cost in latent space using the world model's predictions — explicit model-based planning over abstract states.
+
+## 6. Perception / action / tools
+
+Perception = encoder into latent space; action via the actor/cost modules; designed to be embodiment-agnostic (V-JEPA 2 targets video/robotics).
+
+## 7. Motivation / alignment
+
+A 'cost module' (intrinsic + extrinsic objectives) drives the planner; alignment framed as configurable cost functions; LeCun argues this avoids generative-model pitfalls.
+
+## 8. Self-model / metacognition
+
+The architecture includes the agent's own state in the world model; uncertainty via energy/latent prediction confidence.
+
+## 9. Scalability / path to generality
+
+Generality via hierarchical latent prediction across timescales + a world model that supports planning; V-JEPA 2 is an early instantiation.
+
+## Sources
+
+- LeCun 'A Path Towards Autonomous Machine Intelligence' (JEPA position paper, 2022)
+- V-JEPA 2 (2025)
diff --git a/research/ai_generated_agi_architectures/raw_outputs/mistral.md b/research/ai_generated_agi_architectures/raw_outputs/mistral.md
new file mode 100644
index 0000000..f0e9e7c
--- /dev/null
+++ b/research/ai_generated_agi_architectures/raw_outputs/mistral.md
@@ -0,0 +1,45 @@
+# AGI Architecture Profile — Mistral AI — Small 3.x / open MoE
+**Collection method:** `[reconstructed-from-publications]`
+**Access date:** 2026-07-21
+
+> Reconstructed from the system's published research per the bounty's allowed method; see `sources.md` for primary sources.
+
+## 1. Core paradigm
+
+Efficient open (and commercial) transformers/MoE with strong tool use; no published AGI cognitive architecture. Paradigm: efficient, deployable, agentic models.
+
+## 2. World model / knowledge
+
+Parametric knowledge in weights; no explicit world model/KG.
+
+## 3. Memory systems
+
+Context window working memory; agentic memory via tools/retrieval.
+
+## 4. Learning mechanisms
+
+Pretraining + instruction tuning + RLHF; efficient architectures (MoE, sliding-window attention).
+
+## 5. Reasoning & planning
+
+Tool-augmented reasoning; agentic function calling; less emphasis on long RL reasoning than o/R1.
+
+## 6. Perception / action / tools
+
+Multimodal (Small 3.1 is natively multimodal); strong tool/function calling and on-device deployment.
+
+## 7. Motivation / alignment
+
+RLHF/instruction alignment; objective-following.
+
+## 8. Self-model / metacognition
+
+Not a published architectural component.
+
+## 9. Scalability / path to generality
+
+Generality via efficient scaling + multimodality + agentic tool use + broad deployment.
+
+## Sources
+
+- Mistral Small 3.1 release (2025)
diff --git a/research/ai_generated_agi_architectures/raw_outputs/openai_gpt_o.md b/research/ai_generated_agi_architectures/raw_outputs/openai_gpt_o.md
new file mode 100644
index 0000000..63e934b
--- /dev/null
+++ b/research/ai_generated_agi_architectures/raw_outputs/openai_gpt_o.md
@@ -0,0 +1,46 @@
+# AGI Architecture Profile — OpenAI — GPT / o-series
+**Collection method:** `[reconstructed-from-publications]`
+**Access date:** 2026-07-21
+
+> Reconstructed from the system's published research per the bounty's allowed method; see `sources.md` for primary sources.
+
+## 1. Core paradigm
+
+Scale a transformer language model, then train reasoning with reinforcement learning on chain-of-thought (the o-series 'deliberative alignment' recipe). No named cognitive architecture; the bet is scale + RL-on-reasoning.
+
+## 2. World model / knowledge
+
+Knowledge is implicit in transformer weights; no explicit world model or KG. o-series emits long internal reasoning traces as a scratchpad.
+
+## 3. Memory systems
+
+Working memory = context window. No native persistent episodic/semantic store (added externally via retrieval/tools).
+
+## 4. Learning mechanisms
+
+Pretraining (self-supervised next-token) + RLHF + RL on verifiable rewards for reasoning (GRPO-style). In-context learning at inference; no continual weight updates in deployment.
+
+## 5. Reasoning & planning
+
+Test-time compute: long RL-trained chain-of-thought 'thinking' before answering; tool use for planning. Search is implicit in the learned reasoning policy rather than explicit tree search.
+
+## 6. Perception / action / tools
+
+Multimodal (text/image/audio in, text out); strong tool/function calling and code execution as the action interface.
+
+## 7. Motivation / alignment
+
+RLHF + 'deliberative alignment' (model reasons about its instructions/values explicitly). Objectives set by trainers; corrigibility via training, not architectural.
+
+## 8. Self-model / metacognition
+
+Limited explicit self-model; some calibrated confidence via sampling; reasoning traces give partial introspection.
+
+## 9. Scalability / path to generality
+
+Generality via scaling laws + RL on ever-harder reasoning tasks; capability grows with parameters/data/test-time compute.
+
+## Sources
+
+- OpenAI o-series / deliberative alignment (2024-25)
+- Sam Altman 'Reflections' blog (2025)
diff --git a/research/ai_generated_agi_architectures/raw_outputs/xai_grok.md b/research/ai_generated_agi_architectures/raw_outputs/xai_grok.md
new file mode 100644
index 0000000..1f79019
--- /dev/null
+++ b/research/ai_generated_agi_architectures/raw_outputs/xai_grok.md
@@ -0,0 +1,46 @@
+# AGI Architecture Profile — xAI — Grok
+**Collection method:** `[reconstructed-from-publications]`
+**Access date:** 2026-07-21
+
+> Reconstructed from the system's published research per the bounty's allowed method; see `sources.md` for primary sources.
+
+## 1. Core paradigm
+
+Large (mixture-of-experts) transformer trained to 'understand the universe', with a stated emphasis on curiosity and real-time knowledge. No formal cognitive architecture published.
+
+## 2. World model / knowledge
+
+Parametric knowledge in weights, augmented by real-time access to X (Twitter) data for up-to-date world knowledge.
+
+## 3. Memory systems
+
+Context window working memory; real-time data retrieval as an external memory proxy.
+
+## 4. Learning mechanisms
+
+Large-scale pretraining + RLHF; MoE for efficient scaling.
+
+## 5. Reasoning & planning
+
+Chain-of-thought reasoning modes; tool use; less emphasis on explicit search.
+
+## 6. Perception / action / tools
+
+Multimodal (text/image); tool/function calling; real-time web/X access as tools.
+
+## 7. Motivation / alignment
+
+RLHF alignment; stated goal of truth-seeking/curiosity; 'maximum truth-seeking' framing.
+
+## 8. Self-model / metacognition
+
+Not a published architectural component.
+
+## 9. Scalability / path to generality
+
+Generality via scaling MoE + real-time knowledge integration + reasoning.
+
+## Sources
+
+- xAI Grok system card (2025)
+- xAI 'understand the universe' framing
diff --git a/research/ai_generated_agi_architectures/sources.md b/research/ai_generated_agi_architectures/sources.md
new file mode 100644
index 0000000..cded0aa
--- /dev/null
+++ b/research/ai_generated_agi_architectures/sources.md
@@ -0,0 +1,16 @@
+# Sources
+
+All accessed 2026-07-21.
+
+| System | Provider | Method |
+|--------|----------|--------|
+| Clusy / Claude-family | Anthropic | self-report |
+| GPT/o-series | OpenAI | reconstructed |
+| Gemini/Alpha | Google DeepMind | lab-stated ontology |
+| Grok | xAI | reconstructed |
+| V3/R1 | DeepSeek | reconstructed |
+| Qwen3 | Alibaba | reconstructed |
+| Llama + V-JEPA | Meta | lab-stated architecture |
+| Small 3.x | Mistral | reconstructed |
+
+JEPA is the only genuine named cognitive architecture. DeepMind Levels is a capability ontology, not architecture. Others reconstructed from publications.
\ No newline at end of file
diff --git a/research/ai_generated_agi_architectures/summary.md b/research/ai_generated_agi_architectures/summary.md
new file mode 100644
index 0000000..252be81
--- /dev/null
+++ b/research/ai_generated_agi_architectures/summary.md
@@ -0,0 +1,16 @@
+# Summary
+
+## Common patterns
+- All center on large neural model as core substrate
+- Reasoning = learned CoT (OpenAI o, DeepSeek R1) OR explicit search (DeepMind, PPGW, JEPA)
+- Memory beyond context window only in JEPA + PPGW
+- Alignment = RLHF universally; only JEPA + PPGW embed it architecturally
+
+## Key disagreement
+Token-prediction scaling (OpenAI, DeepSeek, Qwen, xAI, Mistral) vs latent-prediction world models (JEPA, DeepMind, PPGW)
+
+## Notable ideas
+- JEPA latent prediction (model representations, not sensory detail)
+- DeepMind model-based RL + tree search
+- DeepSeek R1 / OpenAI o RL-trained deliberation
+- PPGW global workspace + active inference + corrigibility
\ No newline at end of file
diff --git a/research/ai_generated_agi_architectures/synthesis.md b/research/ai_generated_agi_architectures/synthesis.md
new file mode 100644
index 0000000..f1a9aa4
--- /dev/null
+++ b/research/ai_generated_agi_architectures/synthesis.md
@@ -0,0 +1,17 @@
+# Synthesis: Combined AGI Architecture
+
+| Component | Source |
+|-----------|-------|
+| Latent-prediction world model | Meta JEPA |
+| Model-based RL + tree search | DeepMind Alpha |
+| RL-trained deliberation | DeepSeek R1 / OpenAI o |
+| Global workspace | Clusy PPGW |
+| Episodic/semantic/procedural memory | PPGW + JEPA |
+| Active inference | PPGW |
+| Hierarchical goals | PPGW + DeepMind |
+| Constitutional alignment + corrigibility | PPGW + OpenAI |
+| Self-model + metacognition | PPGW + DeepMind |
+
+**Flow:** Senses → JEPA latent encoder → World model → Global workspace → Planner (latent search + RL deliberation) → Skills → Tools → Prediction error → Update model + memory.
+
+Layered alignment: constitutional priors + value uncertainty + corrigibility guide reasoning.
\ No newline at end of file