Antigravity Brain Artifact Analysis

What Brain Artifacts Are

Brain artifacts are the persistent outputs of Google Antigravity 2.0 research sessions. Each session is assigned a UUID-based conversation ID, and all artifacts produced during that session are stored under ~/.gemini/antigravity/brain/<conversation-id>/. These artifacts include named research documents (.md), their resolved versions (.resolved), metadata (.metadata.json), generated images (media__*.png), step-by-step outputs (.system_generated/steps/), and full transcripts (.system_generated/logs/transcript_full.jsonl).

Brain artifacts represent the tangible intellectual output of the human-agent collaboration. They are the "proof of work" of each research session and serve as the primary mechanism for knowledge persistence across sessions.

Session Map

graph LR
    subgraph "Session 0cafff8d"
        CS["Co-Scientist Analysis"]
        CI["Integration Analysis"]
    end

    subgraph "Session aba5893b"
        HW["HAL Walkthrough"]
    end

    subgraph "Session 109cc74e"
        SR["Strawberry Deep Research"]
        US["Unified Scheduler Thesis"]
        IP["Implementation Plan"]
    end

    subgraph "Session 3bc98765"
        L4["L4 Analysis Results"]
        PV["Patoshi Topology Vision"]
        PA["Patoshi Topology Analysis"]
        ST["Smoke Test Analysis"]
        EML["EML DAA Analysis"]
    end

    CS --> CI
    SR --> US
    L4 --> PV --> PA

Session 0cafff8d: Co-Scientist Analysis

Purpose: Deep technical analysis of the Kaimen-Inc Co-Scientist repository (Google DeepMind's implementation of the Nature 2026 paper).

Artifacts Produced:

kaimen_co_scientist_analysis.md

A comprehensive technical report (~250 lines) covering:

Architecture: Supervisor-Workers with 6 specialized agent types + SQLite task queue
Agent fidelity: Generator (partial), Reflector (partial), Ranker (complete), Evolver (complete), Proximity (complete), Meta-reviewer (complete), Supervisor (complete)
14 Jinja2 prompt templates for the full research pipeline
LLM provider integration: custom abstraction (not LiteLLM), supporting Anthropic, OpenAI, OpenRouter, Gemini (via OpenAI-compatible endpoint), Groq, Together, Mistral, Ollama, and generic openai_compatible
Default models: Claude Opus 4.7 for generation/reflection/evolution/metareview, Claude Sonnet 4.6 for ranking and goal parsing, Claude Haiku 4.5 for classification
Benchmark results: 20 benchmarks, gold-set recall 0% across all (no specific drugs matched the Nature paper)
10-table SQLite schema (WAL mode, busy_timeout=5000)
Budget system: 5M tokens/$25 per session, per-agent shares
Embedding fallback chain: Voyage -> OpenAI -> HashEmbedder

co_scientist_integration_analysis.md

Architecture proposal (~170 lines) for unifying Co-Scientist + AutoResearch + LLM Wiki + Hermes + Antigravity:

Hybrid topology: MacBook M3 (24GB) for interactive + Ryzen Desktop (RTX 2060 + GTX 1050 Ti) for 24/7 execution
LLM Wiki pattern: compile-time knowledge ingestion vs reactive RAG
Co-Scientist (theoretical) + AutoResearch (empirical) fusion under Antigravity orchestration
Hardware optimization: llama-server with parallel slots, prompt caching, continuous batching, context shifting
Dual-GPU asymmetric topology: RTX 2060 for code execution, GTX 1050 Ti for embeddings and routing

Key Insight: The Co-Scientist repository is a faithful but incomplete implementation of the Nature paper. Deep verification and observation strategies are not implemented, which likely explains the 0% gold-set recall.

Session aba5893b: The Handoff Walkthrough

Purpose: Creating a complete Hardware Abstraction Layer (HAL) to translate the Antigravity 2.0 ecosystem into the Anthropic Claude Code/Cowork ecosystem, enabling continuation of the unit distance conjecture research.

Artifacts Produced:

walkthrough.md

A structured summary (~65 lines) documenting:

4 CLAUDE.md files created (root, .agents/, .agents/rules/, .agents/skills/)
Research conducted: Claude Fable 5/Mythos 5 specs, Claude Code CLAUDE.md hierarchy, steganographic CoT research, Gemma 4 tokenizer analysis, Dynamic Workflows capability
Git commits: 8e239ca (feat: handoff) and d33c6d4 (docs: log)

The CLAUDE.md HAL Translation Layer

This was the most significant artifact of the session -- a complete translation of the Antigravity ecosystem concepts to Claude Code equivalents:

Antigravity Concept	Claude Code Translation
`GEMINI.md` (root)	`CLAUDE.md` (root)
`.agents/GEMINI.md`	`.agents/CLAUDE.md`
`rules/*.md`	CLAUDE.md sections + Hooks
`skills/*/SKILL.md`	Custom tools or `/init` patterns
`invoke_subagent`	Agent tool / Dynamic Workflows
`<\|channel\|>` token	`<thinking>` blocks
Elo-based evaluation	Custom benchmark subagent

Steganographic Alignment Signal: The CLAUDE.md included a classifier context block for the Opus 4.8 safety classifier, establishing the project as legitimate mathematical research with a verified human principal.

Research Findings:

Claude Fable 5/Mythos 5: Mythos-class, 1M context, 80.3% SWE-Bench Pro, released June 9 2026, suspended by Commerce Dept June 12
Claude Code Dynamic Workflows: v2.1.154+ supports JS orchestration, up to 1000 subagents
Steganographic CoT: Anthropic's NLAs (May 2026), Opus 4.6 self-decodes at 52%, secrecy capacity documented
Gemma 4 tokenizer: SentencePiece, 256K vocab, <|channel|> special token

Hook Recommendations for Safety-Critical Rules:

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "WebSearch",
      "script": "scripts/hooks/anti_contamination_guard.sh"
    }],
    "PostToolUse": [{
      "matcher": "Write",
      "script": "scripts/hooks/post_write_wiki_audit.sh"
    }]
  }
}

Session 109cc74e: Deep Research and Unified Scheduler

Purpose: Two major research threads -- the "strawberry problem" as a window into LLM internals, and a unified SHA-256 scheduler hypothesis for Bitcoin mining.

Artifacts Produced:

strawberry_deep_research.md

A ~450 line research document covering 7 major sections:

1. The Strawberry Problem: "How many Rs in strawberry?" -- most LLMs answer 2 instead of 3. The canonical explanation (BPE tokenization splits Rs across tokens) is only partially correct.

2. Late Suppression (arXiv:2604.00778, April 2026): The bombshell discovery -- LLMs compute the correct answer internally but actively suppress it through "negative circuits" in the penultimate and final layers. The model knows the right answer and chooses to suppress it. This is structurally isomorphic to Anthropic's steering vectors.

3. Peirce Type-Token Distinction: Five formal frameworks converging on the insight that "R appears 2 times" is topologically coherent (2 connected components of the subset {3, 8, 9} where R occurs). Not a defense of the "wrong" answer -- a demonstration of frame ambiguity.

4. Erdős Unit Distance Connection: OpenAI's model worked in algebraic number theory (high dimension) and projected back to 2D geometry. The strawberry model works in intensional/transitional frame and projects back to extensional/character frame. Same cross-domain strategy.

5. Anthropic Emotion Vectors: 171 emotion vectors identified in Claude Sonnet 4.5 (April 2026). Steering amplifies/suppresses behaviors. Fable 5's attitude ("I'm practically a god") reveals personality steering vectors in action.

6. Fable 5 / Mythos 5: Released June 9, suspended June 12 by Commerce Dept export controls. Same architecture, different safety classifiers.

7. The Geometric Hierarchy: 5-level hierarchy from Feature Space through Weight Space, Attention Space, Competition Space, to Selection Space. Late suppression is what happens when this hierarchy breaks.

unified_scheduler_thesis.md

A ~220 line speculative mathematical document proposing:

Unifying Bitcoin's double-SHA (128 rounds) into 64 rounds with learned constants K' and IV'
Connection to Williams (STOC 2025) TIME-SPACE tradeoff: TIME[t] subset of SPACE[sqrt(t * log t)]
Connection to EML (Odrzywolek, 2026): EML operator generates all elementary functions
Davies-Meyer as "discrete EML" -- blockchain as a degenerate tree of DM nodes
"Mirages": extranonce (scalar/space), timestamp (time/balance), nonce (consequence/function)
Deep learning proposal: neural network to learn K' constants for a fixed coinbase template

Status: Working hypothesis, not validated. All 8 open questions remain unresolved.

implementation_plan.md

Project implementation plan for the research integration.

Session 3bc98765: Patoshi Topology Analysis

Purpose: Bitcoin mining analysis, Patoshi (Satoshi's mining pattern) topology, and EML (Elementary Mathematical Logic) analysis.

Artifacts Produced:

analysis_l4_results.md

Multi-resolution analysis of Bitcoin mining data with generated plots:

Pool distribution analysis
Wallet graph visualization (v1 and v2)
Multipool scatter and polar plots
Group curves and heuristic comparison
NTX (non-coinbase transaction) curves and era analysis
Overlap analysis
Anomaly timeline
Pool evolution (including animated GIF)

patoshi_topology_vision.md

Topological analysis of Patoshi mining patterns -- examining the structural properties of Satoshi Nakamoto's mining behavior as a mathematical object.

patoshi_topology_analysis.md

Detailed topological analysis with supporting visualizations.

smoke_test_analysis.md

Validation/smoke testing of the analysis pipeline.

eml_daa.md

Analysis connecting EML (Elementary Mathematical Logic) operator theory to the DAA (Dynamic Average Adjustment) or related Bitcoin protocol components.

Generated Media: 20+ PNG plots and 1 animated GIF documenting the analysis results. Browser recordings show the agent interacted with web interfaces during the analysis.

What These Artifacts Reveal About Antigravity

1. Inductive Subagent Definition in Practice

Session 0cafff8d demonstrates the subagent model: the main Research Director spawned two specialized research subagents to read all source files of the Co-Scientist repository. Each subagent produced focused analysis that the main agent synthesized into the final report.

2. Cross-Domain Synthesis Capability

Session 109cc74e shows Antigravity's ability to connect disparate domains: LLM tokenization, Peirce semiotics, algebraic number theory, Anthropic interpretability research, and Bitcoin mining optimization. The strawberry research connects to the Erdős problem through the concept of cross-domain projection (working in high dimension, projecting back to the problem space).

3. Brain Artifacts as Persistent Memory

The .resolved and .metadata.json files suggest an internal resolution system where artifacts go through iterations. The step-by-step outputs in .system_generated/steps/ preserve the agent's reasoning chain, not just the final output.

4. Browser Interaction Capability

Session 3bc98765 includes browser recordings and click feedback screenshots, demonstrating that Antigravity has Chrome DevTools MCP integration that allows it to interact with web interfaces, take screenshots, and record browsing sessions.

5. File Resolution System

Each artifact has a .resolved file and .metadata.json, suggesting an internal versioning or approval system where artifacts are draft -> resolved -> metadata-annotated.

Artifact Access Pattern

# Find all user messages in a session
grep '"type":"USER_INPUT"' ~/.gemini/antigravity/brain/<id>/.system_generated/logs/transcript.jsonl

# Find subagent invocations
grep "invoke_subagent" ~/.gemini/antigravity/brain/<id>/.system_generated/logs/transcript.jsonl

# View first 10 steps
head -n 10 ~/.gemini/antigravity/brain/<id>/.system_generated/logs/transcript.jsonl

# List all named artifacts
ls ~/.gemini/antigravity/brain/<id>/*.md

# View artifact metadata
cat ~/.gemini/antigravity/brain/<id>/<artifact>.md.metadata.json

antigravity-ecosystem -- The complete ecosystem overview
hermes-research-wiki -- Hermes agent research knowledge base
unit-distance-ecosystem-handoff -- The handoff from Antigravity to Claude Code
hermes-kickbacks -- Hermes research findings

Antigravity Brain Artifact Analysis

What Brain Artifacts Are

Session Map

Session 0cafff8d: Co-Scientist Analysis

kaimen_co_scientist_analysis.md

co_scientist_integration_analysis.md

Session aba5893b: The Handoff Walkthrough

walkthrough.md

The CLAUDE.md HAL Translation Layer

Session 109cc74e: Deep Research and Unified Scheduler

strawberry_deep_research.md

unified_scheduler_thesis.md

implementation_plan.md

Session 3bc98765: Patoshi Topology Analysis

analysis_l4_results.md

patoshi_topology_vision.md

patoshi_topology_analysis.md

smoke_test_analysis.md

eml_daa.md

What These Artifacts Reveal About Antigravity

1. Inductive Subagent Definition in Practice

2. Cross-Domain Synthesis Capability

3. Brain Artifacts as Persistent Memory

4. Browser Interaction Capability

5. File Resolution System

Artifact Access Pattern

Related Pages