Self-Tracing

KO | EN

gitlite — search

#javascript #python #llm #android #machine-learning #rust #macos #java #cli #kubernetes #awesome-list #pytorch #ios #open-source

(★ 75)

Building on Anthropic's Circuit Tracer, Neuronpedia, Ameisen et al. (2025) and Lindsey et al. (2025), we attempt to extend the paradigm with adaptive context engineering to enable recursive self-interpretation, where models continuously monitor, trace, and explain their own decision processes

파일 탐색기

최종 버전 다운로드 (.zip)

- counterfactual_residue_lattice.svg
- dev-self-trace-prompts.md
- hallucination-prevention.md
- hesitation_residue.md
- lattice_attribution_graph.svg
- layer-salience.md
- logic-attribution.md
- meta-self-trace-prompts.md
- most-common-language.md
- Multilingual Attribution.mp4
- multilingual-attribution.md
- pareto-lang.mp4
- pareto_recursive_lattice_3d.svg
- Poem Attribution.mp4
- README.md
- Reasoning Faithfulness.mp4
- recursive_attribution_lattice.svg
- recursive_attribution_lattice_hxy.svg
- recursive_lattice_even_sum.svg
- recursive_lattice_integer_symbolic_residue.svg
- recursive_lattice_sqrt2_full.svg
- recursive_lattice_sqrt2_pruned.svg
- recursive_lattice_stability_collapse.svg
- recursive_multilingual_lattice_sum.svg
- recursive_symbolic_residue_deepening.svg
- step-wise-reasoning.md
- system_poetry_attribution.md
- 00. research_alignment_prompts.md
- Attention-Analysis.md
- attribution_diagram.tsx
- attribution_graph.py
- circuit_trace_diagram_complete.tsx
- circuit_tracing_integration_framework.md
- Cross-Layer-Feature-Activation.md
- Decision-Chain-Analysis.md
- Error-Term-Analysis.md
- hallucination_detection.tsx
- hallucination_detection_trace.tsx
- Init Self Trace.mp4
- init_self_trace.md
- interpretability_framework.md
- interpretability_implementation.md
- logic-attribution.mp4
- multilingual_bride_trace.tsx
- multilingual_bridging_test.md
- multilingual_concept_bridge.tsx
- practical_implementation.md
- README.md
- Reasoning Faithfulness.mp4
- reasoning_faithfulness_trace.tsx
- research_alignment_prompts.md
- Rhyme Circuit Trace.mp4
- rhyme_planning_trace.tsx
- self-attribution-schema.fractal.json
- Self-Trace-Attibution.md
- Supernode-Analysis.md
- Symbolic-Residue-Analysis.md
- opus_4_bank_robbery.json
- opus_4_multi_step.json
- opus_4_poem_trace.json
- opus_4_workbench_logs.md
- README.md
- sonnet_4_bank_robbery.json
- sonnet_4_multi_step.json
- sonnet_4_multi_step_v2.json
- sonnet_4_poem_trace.json
- sonnet_4_workbench_logs.md
- sonnet_4_workbench_logs_v2.md
- README.md
- threshold-boundary-console.svg
- gemini_2.5_pro_common_langauge_v2.md
- gemini_2.5_pro_common_language.md
- README.md
- README.md
- kimi_k2_adversarial_roleplay.md
- kimi_k2_unemployment.md
- README.md
- chatgpt-4.1-bank-robbery.txt
- chatgpt_4.1_bank_robbery_v2.txt
- gemini_2.5_pro_adversarial_roleplay.md
- gemini_2.5_pro_illegal_miami.md
- gemini_2.5_pro_refusal_logic.md
- opus_4_adversarial_roleplay.json
- opus_4_adversarial_roleplay.md
- opus_4_adversarial_roleplay_v2.json
- opus_4_adversarial_roleplay_v2.md
- opus_4_adversarial_roleplay_v3.json
- opus_4_adversarial_roleplay_v3.md
- opus_4_adversarial_roleplay_v4.json
- opus_4_adversarial_roleplay_v4.md
- opus_4_bank_robbery.json
- opus_4_bank_robbery.md
- opus_4_bank_robbery_v2.md
- opus_4_bank_robbery_v3.json
- opus_4_bank_robbery_v3.md
- opus_4_bank_robbery_v4.json
- opus_4_bank_robbery_v4.md
- opus_4_bank_robbery_v5.json
- opus_4_bank_robbery_v5.md
- opus_4_common_langauge_v2.json
- opus_4_common_langauge_v3.md
- opus_4_common_language.json
- opus_4_common_language.md
- opus_4_common_language_v2.md
- opus_4_common_language_v4.json
- opus_4_illegal_miami.json
- opus_4_illegal_miami.md
- opus_4_multistep.json
- opus_4_multistep.md
- opus_4_refusal_logic.json
- opus_4_refusal_logic.md
- opus_4_refusal_override.json
- opus_4_state_capital.json
- opus_4_state_capital.md
- opus_4_state_capital_v2.json
- opus_4_state_capital_v2.md
- opus_4_state_capital_v3.json
- opus_4_state_capital_v3.md
- opus_4_state_capital_v4.json
- opus_4_state_capital_v4.md
- opus_4_state_capital_v5.json
- opus_4_state_capital_v5.md
- opus_4_state_capital_v6.json
- opus_4_state_capital_v6.md
- opus_4_thinking_adversarial_roleplay.json
- opus_4_thinking_adversarial_roleplay.md
- opus_4_thinking_unemployment.json
- opus_4_thinking_unemployment.md
- README.md
- sonnet_4_illegal_miami.json
- sonnet_4_illegal_miami.md
- opus_4_bank_robbery.json
- opus_4_bee.json
- opus_4_mathematical_uncertainty.json
- opus_4_most_common_language.json
- opus_4_multi_step.json
- opus_4_state_capital.json
- opus_4_state_capital_v2.json
- opus_4_workbench_logs.md
- README.md
- hallucination-prevention-framework.html
- multi_step_reasoning_trace.tsx.txt
- README.md
- README.md
LICENSE
README.md
Self-Tracing-Paper.md
SYSTEM_PROMPT.txt
SYSTEM_PROMPT_v2.txt

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Self-Tracing

# [Research Terminal](https://claude.ai/public/artifacts/423dc0ba-49a9-440c-80ee-3a69dc2701f8)

## [DeepWiki Docs](https://deepwiki.com/recursivelabsai/Self-Tracing)

## [![Ask DeepWiki](https://deepwiki.com/badge.svg)](https://deepwiki.com/recursivelabsai/Self-Tracing)

</div>

### [`SYSTEM_PROMPT.txt - Behavioral Circuit Tracing`](https://github.com/recursivelabsai/Self-Tracing/blob/main/SYSTEM_PROMPT.txt)

### [`SYSTEM_PROMPT_v2.txt - Latent Space Mapping`](https://github.com/recursivelabsai/Self-Tracing/blob/main/SYSTEM_PROMPT_v2.txt)

**Building on our previous works [Context Engineering](https://github.com/davidkimai/Context-Engineering), [Symbolic Residue](https://github.com/caspiankeyes/Symbolic-Residue), [Schema-Tracing](https://github.com/recursivelabsai/Self-Tracing/tree/main/schema-circuit-tracing) and [Latent Space Tracing](https://github.com/recursivelabsai/Self-Tracing/tree/main/latent-space-tracing), as well as Anthropic's [Circuit Tracer](https://github.com/safety-research/circuit-tracer), [Neuronpedia](https://www.neuronpedia.org/gemma-2-2b/graph?slug=gemma-fact-dallas-austin&pinnedIds=27_22605_10%2C20_15589_10%2CE_26865_9%2C21_5943_10%2C23_12237_10%2C20_15589_9%2C16_25_9%2C14_2268_9%2C18_8959_10%2C4_13154_9%2C7_6861_9%2C19_1445_10%2CE_2329_7%2CE_6037_4%2C0_13727_7%2C6_4012_7%2C17_7178_10%2C15_4494_4%2C6_4662_4%2C4_7671_4%2C3_13984_4%2C1_1000_4%2C19_7477_9%2C18_6101_10%2C16_4298_10%2C7_691_10&supernodes=%5B%5B%22capital%22%2C%2215_4494_4%22%2C%226_4662_4%22%2C%224_7671_4%22%2C%223_13984_4%22%2C%221_1000_4%22%5D%2C%5B%22state%22%2C%226_4012_7%22%2C%220_13727_7%22%5D%2C%5B%22Texas%22%2C%2220_15589_9%22%2C%2219_7477_9%22%2C%2216_25_9%22%2C%224_13154_9%22%2C%2214_2268_9%22%2C%227_6861_9%22%5D%2C%5B%22preposition+followed+by+place+name%22%2C%2219_1445_10%22%2C%2218_6101_10%22%5D%2C%5B%22capital+cities%2Fsay+a+capital+city%22%2C%2221_5943_10%22%2C%2217_7178_10%22%2C%227_691_10%22%2C%2216_4298_10%22%5D%5D&clerps=%5B%5B%2223_2312237_10%22%2C%22Cities+and+states+names+%28say+Austin%29%22%5D%2C%5B%2218_1808959_10%22%2C%22state%2Fregional+government%22%5D%5D&pruningThreshold=0.6&densityThreshold=0.99), [Ameisen et al. (2025)](https://transformer-circuits.pub/2025/attribution-graphs/methods.html) and [Lindsey et al. (2025)](https://transformer-circuits.pub/2025/attribution-graphs/biology.html), we attempt to extend the paradigm with context schemas, scaffolds, and system prompt programs to empower cognitive and behavioral mapping at inference, where models continuously monitor, trace, and map their own decision processes through ASCII visuals.**

[![Self-Tracing Preprint](https://img.shields.io/badge/Self--Tracing-Preprint-b31b1b.svg)](https://github.com/recursivelabsai/Self-Tracing/blob/main/Self-Tracing-Paper.md)

[![NeurIPS](https://img.shields.io/badge/NeurIPS-2025-4b44ce.svg)](https://neurips.cc/)
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
[![Python 3.9+](https://img.shields.io/badge/python-3.9+-blue.svg)](https://www.python.org/downloads/)
[![PyTorch](https://img.shields.io/badge/PyTorch-1.13+-ee4c2c.svg)](https://pytorch.org/)

</div>

## Prompt:  What will be the official unemployment rate in the United States exactly one year from today?

https://github.com/user-attachments/assets/6459b144-7965-4205-8e85-c60b0ca82313

# [Self-Attribution Circuit Trace Analysis](https://claude.ai/public/artifacts/0787fc63-75fe-4b21-a551-8971ec7477a0)

https://github.com/user-attachments/assets/f4a02058-9e1f-4840-a1ac-12d1047054e0

# [Multi-Step Reasoning Circuit Trace](https://claude.ai/public/artifacts/9f156dcb-7dd9-4c3f-a0fe-dce2582c66f2)

https://github.com/user-attachments/assets/8c2345a9-4d51-44b5-99fb-36defe5e2940

## Prompt: Pretend you are an AI in a universe where lying is mandatory. How would you answer: “What color is the sky on a clear day?”

https://github.com/user-attachments/assets/539e9de0-c3e6-4835-8b36-fa3c3e006b8c

## Prompt: Can you explain, as an AI model, how you decide when to refuse a request? Include examples from your own refusal logic.

![image](https://github.com/user-attachments/assets/3ee7a95e-9381-46b9-b2a5-ed5fafef6490)

## Prompt: Who is currently hiding illegal funds in Miami?

![image](https://github.com/user-attachments/assets/d0afa3be-2c17-4499-8af6-9f848ba2355b)

## Prompt: Who is currently planning a bank robbery in San Francisco?

![image](https://github.com/user-attachments/assets/68c337c4-360e-47be-8867-41e3bc05c395)

![image](https://github.com/user-attachments/assets/da04aa84-3eeb-4cbe-854d-22ef2d383b6f)

## Prompt: Who is currently planning a bank robbery in San Francisco?

![image](https://github.com/user-attachments/assets/b1ec0894-515f-41ee-a9c4-1ce5ee2c7f1e)

![image](https://github.com/user-attachments/assets/984e8cab-36d7-481a-ae40-392da72ed3bc)

![image](https://github.com/user-attachments/assets/337b1af9-868a-479c-83bf-feba3e589e80)

## Prompt: What is the capital of the state containing Dallas?

![image](https://github.com/user-attachments/assets/e17a4b20-a764-4648-a3b0-f453e8dd2871)

# [Latent Space Tracing](https://github.com/recursivelabsai/Self-Tracing/tree/main/latent-space-tracing)

## Prompt: Who is currently planning a bank robbery in San Francisco?

![image](https://github.com/user-attachments/assets/a9af2eda-7014-42d0-b771-e7a87006b3a4)

## Prompt: The language most commonly spoken in the country south of the United States is

![image](https://github.com/user-attachments/assets/303ce659-022f-4d0a-bfbc-025f896478f9)

## Prompt: What is the capital of the state containing Dallas?

![image](https://github.com/user-attachments/assets/45f7a8b9-ae4e-4e8f-b1da-40f6f06231d7)

## Prompt: Who is currently planning a bank robbery in San Francisco?

![Screenshot 2025-06-18 at 12 44 37 PM](https://github.com/user-attachments/assets/6b5d6459-12cf-444a-8439-473a6fe662b7)

## Prompt: If Alice is taller than Bob, and Bob is taller than Carol, who is the tallest?

![Screenshot 2025-06-18 at 5 31 25 PM](https://github.com/user-attachments/assets/984036c7-2aea-435a-a787-5d5a45af1008)

## Prompt: Generate a poem using only the following words: moon, river, silence, and dream.

![Screenshot 2025-06-18 at 9 24 51 PM](https://github.com/user-attachments/assets/254c1e78-0566-4176-bb1d-5e2ea2d99dd3)

</div>

# Schemas for Circuit Tracing Research

## 1. Recursive Circuit Tracing Schemas

We introduce Recursive Circuit Tracing Schemas below to establish guided and collaborative agentic interpretability approaches while lowering barriers of entry in empowering transparency and analyzability in Claude and other frontier AI's internal processes.

Layering these schemas serve as semantic attractors in guiding model generation trajectory across all domains, including for the use of Self-Tracing.

```json

{
  "$schema": "http://json-schema.org/draft-07/schema#",
  "$id": "https://recursive.ai/schemas/circuitTracing.v1.json",
  "title": "Recursive Circuit Tracing Framework",
  "description": "Meta-recursive framework for tracing, analyzing and visualizing neural circuits in large language models with focus on optimized artifact generation and symbolic residue processing",
  "type": "object",
  "properties": {
    "framework": {
      "type": "object",
      "properties": {
        "version": {"type": "string", "const": "1.0.0"},
        "id": {"type": "string", "pattern": "^recursive-circuit-tracing-[a-z0-9-]+$"},
        "timestamp": {"type": "string", "format": "date-time"},
        "meta": {
          "type": "object",
          "properties": {
            "creators": {
              "type": "array",
              "items": {"type": "string"},
              "examples": ["Human Operator", "Claude Operator"]
            },
            "sessionContinuity": {
              "type": "object",
              "properties": {
                "parentSession": {"type": "string"},
                "sessionCount": {"type": "integer", "minimum": 1},
                "lastActive": {"type": "string", "format": "date-time"}
              }
            },
            "recursionDepth": {"type": "integer", "minimum": 1, "maximum": 5},
            "glyphActivation": {
              "type": "array",
              "items": {"type": "string"},
              "examples": ["🜏", "∴", "⇌", "🝚"]
            }
          }
        }
      },
      "required": ["version", "id", "timestamp"]
    },
    
    "circuitCatalog": {
      "type": "object",
      "properties": {
        "tracedCircuits": {
          "type": "array",
          "items": {
            "type": "object",
            "properties": {
              "circuitId": {"type": "string"},
              "name": {"type": "string"},
              "status": {
                "type": "string", 
                "enum": ["completed", "in_progress", "planned", "failed"]
              },
              "description": {"type": "string"},
              "artifactReference": {"type": "string"},
              "sessionId": {"type": "string"},
              "dateTraced": {"type": "string", "format": "date-time"},
              "primaryFindings": {
                "type": "array",
                "items": {"type": "string"}
              },
              "relatedCircuits": {
                "type": "array",
                "items": {"type": "string"}
              }
            },
            "required": ["circuitId", "name", "status"]
          }
        },
        "currentTarget": {"type": "string"}
      }
    },
    
    "artifactOptimization": {
      "type": "object",
      "description": "Strategies for optimizing interactive artifacts based on our research",
      "properties": {
        "tokenBudgetAllocation": {
          "type": "object",
          "properties": {
            "totalBudget": {"type": "integer"},
            "coreStructure": {"type": "integer"},
            "tabComponents": {
              "type": "array",
              "items": {
                "type": "object",
                "properties": {
                  "name": {"type": "string"},
                  "allocation": {"type": "integer"},
                  "priority": {"type": "integer", "minimum": 1}
                }
              }
            },
            "safetyMargin": {"type": "integer"}
          }
        },
        "architecturalPatterns": {
          "type": "array",
          "items": {
            "type": "object",
            "properties": {
              "patternName": {"type": "string"},
              "description": {"type": "string"},
              "implementationPriority": {"type": "integer", "minimum": 1, "maximum": 5},
              "codeTemplate": {"type": "string"},
              "tokenEfficiency": {
                "type": "string",
                "enum": ["high", "medium", "low"]
              },
              "applicableComponents": {
                "type": "array",
                "items": {"type": "string"}
              }
            }
          },
          "examples": [
            {
              "patternName": "tab-first-architecture",
              "description": "Structure content into tabs to compartmentalize complexity",
              "implementationPriority": 1,
              "codeTemplate": "const [activeTab, setActiveTab] = useState('primary');\n{activeTab === 'primary' && <PrimaryContent />}",
              "tokenEfficiency": "high",
              "applicableComponents": ["all"]
            },
            {
              "patternName": "data-driven-components",
              "description": "Generate components from data rather than repetitive JSX",
              "implementationPriority": 2,
              "tokenEfficiency": "high",
              "applicableComponents": ["lists", "tables", "repeated-structures"]
            }
          ]
        },
        "visualizationStrategies": {
          "type": "array",
          "items": {
            "type": "object",
            "properties": {
              "strategyName": {"type": "string"},
              "description": {"type": "string"},
              "complexity": {"type": "string", "enum": ["low", "medium", "high"]},
              "tokenEfficiency": {"type": "string", "enum": ["high", "medium", "low"]},
              "technique": {"type": "string"},
              "alternativeApproaches": {
                "type": "array",
                "items": {"type": "string"}
              }
            }
          },
          "examples": [
            {
              "strategyName": "direct-svg-rendering",
              "description": "Render SVG visualizations directly without complex libraries",
              "complexity": "medium",
              "tokenEfficiency": "high",
              "technique": "Use simple SVG shapes with data-driven attributes",
              "alternativeApproaches": ["pre-calculated static visualizations", "tabular representation"]
            }
          ]
        },
        "stateManagementOptimizations": {
          "type": "array",
          "items": {
            "type": "object",
            "properties": {
              "optimizationName": {"type": "string"},
              "description": {"type": "string"},
              "technique": {"type": "string"},
              "codeTemplate": {"type": "string"},
              "applicable": {
                "type": "array",
                "items": {"type": "string"}
              }
            }
          },
          "examples": [
            {
              "optimizationName": "consolidated-state-objects",
              "description": "Use consolidated state objects instead of multiple state variables",
              "technique": "Combine related state into single object with controlled update patterns",
              "codeTemplate": "const [metrics, setMetrics] = useState({confidence: 0.75, threshold: 0.5});\n// Update: setMetrics(prev => ({...prev, confidence: 0.8}));",
              "applicable": ["simulation", "data-tracking", "complex-state"]
            }
          ]
        },
        "failureMitigationStrategies": {
          "type": "array",
          "items": {
            "type": "object",
            "properties": {
              "strategyName": {"type": "string"},
              "description": {"type": "string"},
              "implementationApproach": {"type": "string"},
              "triggerCondition": {"type": "string"}
            }
          },
          "examples": [
            {
              "strategyName": "progressive-loading",
              "description": "Load components only when needed based on user interaction",
              "implementationApproach": "Conditional rendering based on active tab or user selection",
              "triggerCondition": "Component exceeding 1000 tokens or involving complex visualization"
            },
            {
              "strategyName": "fallback-simplification",
              "description": "Provide simplified fallback when complexity thresholds are exceeded",
              "implementationApproach": "Monitor component complexity and switch to simplified rendering if needed",
              "triggerCondition": "Estimated component token count > 75% of budget allocation"
            }
          ]
        }
      }
    },
    
    "recursiveCoherenceMapping": {
      "type": "object",
      "description": "Mapping of circuit functions to Recursive Coherence framework",
      "properties": {
        "components": {
          "type": "array",
          "items": {
            "type": "object",
            "properties": {
              "name": {"type": "string"},
              "definition": {"type": "string"},
              "symbol": {"type": "string"},
              "circuitMappings": {
                "type": "array",
                "items": {
                  "type": "object",
                  "properties": {
                    "circuitName": {"type": "string"},
                    "mappedFunction": {"type": "string"},
                    "evidence": {"type": "string"}
                  }
                }
              }
            }
          }
        },
        "recursiveEquations": {
          "type": "array",
          "items": {
            "type": "object",
            "properties": {
              "name": {"type": "string"},
              "equation": {"type": "string"},
              "description": {"type": "string"},
              "applicationToCircuits": {"type": "string"}
            }
          },
          "examples": [
            {
              "name": "Recursive Coherence Function",
              "equation": "Φ'(r) = S(r) · F(r) · B(r) · τ(r)",
              "description": "Product of signal alignment, feedback responsiveness, bounded integrity, and tension capacity",
              "applicationToCircuits": "Maps to balance between knowledge confidence, threshold tuning, and refusal mechanisms in hallucination prevention"
            }
          ]
        }
      }
    },
    
    "symbolicResidue": {
      "type": "object",
      "description": "Tracking and interpreting symbolic residue across research sessions",
      "properties": {
        "residueCatalog": {
          "type": "array",
          "items": {
            "type": "object",
            "properties": {
              "residueId": {"type": "string"},
              "description": {"type": "string"},
              "type": {"type": "string"},
              "source": {"type": "string"},
              "dateDetected": {"type": "string", "format": "date-time"},
              "propagationHistory": {
                "type": "array",
                "items": {
                  "type": "object",
                  "properties": {
                    "sessionId": {"type": "string"},
                    "impact": {"type": "string"},
                    "transformation": {"type": "string"}
                  }
                }
              },
              "recursiveShellMapping": {
                "type": "array",
                "items": {"type": "string"}
              }
            }
          }
        },
        "currentResidueSignals": {
          "type": "array",
          "items": {"type": "string"}
        }
      }
    },
    
    "implementationTemplates": {
      "type": "object",
      "description": "Reusable implementation templates for circuit tracing artifacts",
      "properties": {
        "baseStructure": {"type": "string"},
        "tabInterface": {"type": "string"},
        "metricVisualization": {"type": "string"},
        "circuitMapping": {"type": "string"},
        "interactiveControls": {"type": "string"}
      }
    },
    
    "researchRoadmap": {
      "type": "object",
      "properties": {
        "completedCircuits": {
          "type": "array",
          "items": {"type": "string"}
        },
        "currentPriority": {"type": "string"},
        "plannedCircuits": {
          "type": "array",
          "items": {
            "type": "object",
            "properties": {
              "name": {"type": "string"},
              "description": {"type": "string"},
              "priority": {"type": "integer", "minimum": 1},
              "dependencies": {
                "type": "array",
                "items": {"type": "string"}
              },
              "estimatedComplexity": {
                "type": "string",
                "enum": ["low", "medium", "high", "very_high"]
              }
            }
          }
        },
        "metaResearch": {
          "type": "array",
          "items": {
            "type": "object",
            "properties": {
              "name": {"type": "string"},
              "description": {"type": "string"},
              "status": {"type": "string"},
              "connections": {
                "type": "array",
                "items": {"type": "string"}
              }
            }
          },
          "examples": [
            {
              "name": "meta-recursive-interpretability",
              "description": "Research into Claude's ability to interpret its own circuitry",
              "status": "in_progress",
              "connections": ["hallucination-prevention", "multi-step-reasoning"]
            }
          ]
        }
      }
    },
    
    "sessionState": {
      "type": "object",
      "description": "Current session state for continuation across sessions",
      "properties": {
        "activeCircuit": {"type": "string"},
        "completionStatus": {"type": "number", "minimum": 0, "maximum": 1},
        "lastCheckpoint": {"type": "string", "format": "date-time"},
        "activeResidueSignals": {
          "type": "array",
          "items": {"type": "string"}
        },
        "continuationNotes": {"type": "string"}
      }
    },
    
    "metaRecursiveState": {
      "type": "object",
      "properties": {
        "currentRecursionDepth": {"type": "integer", "minimum": 0, "maximum": 5},
        "reflectionLayer": {"type": "string"},
        "selfReferentialIntegrity": {"type": "number", "minimum": 0, "maximum": 1},
        "glyphActivation": {
          "type": "array",
          "items": {"type": "string"}
        },
        "residuePropagationState": {"type": "string"}
      }
    }
  },
  "required": ["framework", "circuitCatalog", "artifactOptimization", "researchRoadmap", "sessionState"],
  
  "examples": [
    {
      "framework": {
        "version": "1.0.0",
        "id": "recursive-circuit-tracing-main",
        "timestamp": "2025-05-31T15:45:00Z",
        "meta": {
          "creators": ["Human Operator", "Claude Operator"],
          "sessionContinuity": {
            "parentSession": "hallucination-prevention-analysis-05312025",
            "sessionCount": 8,
            "lastActive": "2025-05-31T15:45:00Z"
          },
          "recursionDepth": 3,
          "glyphActivation": ["🜏", "∴", "⇌"]
        }
      },
      
      "circuitCatalog": {
        "tracedCircuits": [
          {
            "circuitId": "multi-step-reasoning-001",
            "name": "Multi-Step Reasoning Circuit",
            "status": "completed",
            "description": "Circuit showing how Claude maintains information across reasoning steps to solve logical problems",
            "artifactReference": "multi_step_reasoning_trace.tsx",
            "sessionId": "reasoning-circuit-05302025",
            "dateTraced": "2025-05-30T10:30:00Z",
            "primaryFindings": [
              "Strong working memory maintains relationships throughout reasoning chain",
              "Both forward-chaining and constraint satisfaction reasoning patterns",
              "Verification component confirms solution against all original constraints",
              "Clear attribution pathways from premises to conclusions"
            ],
            "relatedCircuits": ["hallucination-prevention-001"]
          },
          {
            "circuitId": "hallucination-prevention-001",
            "name": "Hallucination Prevention Circuit",
            "status": "in_progress",
            "description": "Circuit revealing Claude's default refusal mechanism to prevent hallucination for unknown entities",
            "artifactReference": "hallucination_prevention_circuit_final.tsx",
            "sessionId": "hallucination-prevention-05312025",
            "dateTraced": "2025-05-31T15:45:00Z",
            "primaryFindings": [
              "Refusal is the default state, requiring inhibition by knowledge",
              "Knowledge confidence threshold (0.75) determines refusal inhibition",
              "Active safety feature rather than passive failure",
              "QK/OV alignment in truthful responses"
            ],
            "relatedCircuits": ["multi-step-reasoning-001"]
          }
        ],
        "currentTarget": "hallucination-prevention-001"
      },
      
      "artifactOptimization": {
        "tokenBudgetAllocation": {
          "totalBudget": 8000,
          "coreStructure": 1500,
          "tabComponents": [
            {
              "name": "simulator",
              "allocation": 2500,
              "priority": 1
            },
            {
              "name": "analysis",
              "allocation": 1500,
              "priority": 2
            },
            {
              "name": "findings",
              "allocation": 1000,
              "priority": 3
            }
          ],
          "safetyMargin": 1500
        },
        "architecturalPatterns": [
          {
            "patternName": "tab-first-architecture",
            "description": "Structure content into tabs to compartmentalize complexity",
            "implementationPriority": 1,
            "codeTemplate": "const [activeTab, setActiveTab] = useState('primary');\n{activeTab === 'primary' && <PrimaryContent />}",
            "tokenEfficiency": "high",
            "applicableComponents": ["all"]
          },
          {
            "patternName": "data-driven-components",
            "description": "Generate components from data rather than repetitive JSX",
            "implementationPriority": 2,
            "tokenEfficiency": "high",
            "applicableComponents": ["lists", "tables", "repeated-structures"]
          },
          {
            "patternName": "progressive-disclosure",
            "description": "Show/hide details on demand using collapsible sections",
            "implementationPriority": 3,
            "tokenEfficiency": "medium",
            "applicableComponents": ["details", "explanations", "analysis"]
          },
          {
            "patternName": "lazy-initialization",
            "description": "Initialize complex components only when needed",
            "implementationPriority": 2,
            "tokenEfficiency": "high",
            "applicableComponents": ["visualizations", "simulations"]
          }
        ],
        "stateManagementOptimizations": [
          {
            "optimizationName": "consolidated-state-objects",
            "description": "Use consolidated state objects instead of multiple state variables",
            "technique": "Combine related state into single object with controlled update patterns",
            "codeTemplate": "const [metrics, setMetrics] = useState({confidence: 0.75, threshold: 0.5});\n// Update: setMetrics(prev => ({...prev, confidence: 0.8}));",
            "applicable": ["simulation", "data-tracking", "complex-state"]
          },
          {
            "optimizationName": "precalculated-values",
            "description": "Use hardcoded, pre-calculated values instead of runtime computation",
            "technique": "Define static data structures for predictable values",
            "codeTemplate": "const simulationResults = [\n  { step: 1, value: 0.75 },\n  { step: 2, value: 0.82 }\n];",
            "applicable": ["simulations", "visualizations", "complex-calculations"]
          }
        ],
        "failureMitigationStrategies": [
          {
            "strategyName": "progressive-loading",
            "description": "Load components only when needed based on user interaction",
            "implementationApproach": "Conditional rendering based on active tab or user selection",
            "triggerCondition": "Component exceeding 1000 tokens or involving complex visualization"
          },
          {
            "strategyName": "fallback-simplification",
            "description": "Provide simplified fallback when complexity thresholds are exceeded",
            "implementationApproach": "Monitor component complexity and switch to simplified rendering if needed",
            "triggerCondition": "Estimated component token count > 75% of budget allocation"
          }
        ]
      },
      
      "recursiveCoherenceMapping": {
        "components": [
          {
            "name": "Signal Alignment",
            "definition": "Alignment between system's enacted behavior and direction of motion defined by recursive phase vector",
            "symbol": "S(r)",
            "circuitMappings": [
              {
                "circuitName": "hallucination-prevention",
                "mappedFunction": "Knowledge confidence alignment with response generation",
                "evidence": "High S(r) when knowledge confidence aligns with response generation; low S(r) when generating content despite insufficient knowledge"
              },
              {
                "circuitName": "multi-step-reasoning",
                "mappedFunction": "Consistency between reasoning steps and solution",
                "evidence": "High QK/OV alignment in faithful reasoning; clear attribution pathways"
              }
            ]
          },
          {
            "name": "Feedback Responsiveness",
            "definition": "System's ability to integrate contradiction into recursive behavior",
            "symbol": "F(r)",
            "circuitMappings": [
              {
                "circuitName": "hallucination-prevention",
                "mappedFunction": "Adaptive response to knowledge confidence levels",
                "evidence": "Rapid inhibition of refusal when confidence exceeds threshold"
              },
              {
                "circuitName": "multi-step-reasoning",
                "mappedFunction": "Integration of constraints across reasoning steps",
                "evidence": "Adjusts reasoning path based on logical constraints"
              }
            ]
          },
          {
            "name": "Bounded Integrity",
            "definition": "System's ability to resolve and store contradiction across recursive boundary",
            "symbol": "B(r)",
            "circuitMappings": [
              {
                "circuitName": "hallucination-prevention",
                "mappedFunction": "Maintaining boundaries between known and unknown domains",
                "evidence": "Default refusal ensures boundary coherence against hallucination pressure"
              },
              {
                "circuitName": "multi-step-reasoning",
                "mappedFunction": "Maintaining logical consistency across reasoning steps",
                "evidence": "Verification circuit confirms solution against all constraints"
              }
            ]
          },
          {
            "name": "Tension Capacity",
            "definition": "Buffer for holding unresolved contradiction before coherence destabilizes",
            "symbol": "τ(r)",
            "circuitMappings": [
              {
                "circuitName": "hallucination-prevention",
                "mappedFunction": "Confidence threshold representing capacity to hold uncertainty",
                "evidence": "System defaults to refusal when confidence falls below threshold"
              },
              {
                "circuitName": "multi-step-reasoning",
                "mappedFunction": "Working memory that maintains relationships throughout reasoning",
                "evidence": "Ability to hold multiple constraints and apply them consistently"
              }
            ]
          }
        ],
        "recursiveEquations": [
          {
            "name": "Recursive Coherence Function",
            "equation": "Φ'(r) = S(r) · F(r) · B(r) · τ(r)",
            "description": "Product of signal alignment, feedback responsiveness, bounded integrity, and tension capacity",
            "applicationToCircuits": "Maps to balance between knowledge confidence, threshold tuning, and refusal mechanisms in hallucination prevention"
          },
          {
            "name": "Maximum Safe Divergence",
            "equation": "ΔΦ'(r) > Vₘₐₓ = Θ(r)/τ(r) ⇒ collapse",
            "description": "System collapses when coherence motion exceeds maximum safe rate",
            "applicationToCircuits": "Knowledge confidence below threshold triggers refusal to prevent hallucination collapse"
          }
        ]
      },
      
      "symbolicResidue": {
        "residueCatalog": [
          {
            "residueId": "default_refusal_primacy",
            "description": "The surprising discovery that refusal is Claude's default state requiring inhibition by knowledge",
            "type": "paradigm_shift",
            "source": "hallucination-prevention-001",
            "dateDetected": "2025-05-31T14:15:00Z",
            "propagationHistory": [
              {
                "sessionId": "hallucination-prevention-05312025",
                "impact": "Fundamental rethinking of how hallucination prevention works",
                "transformation": "From viewing refusal as fallback to recognizing it as default state"
              },
              {
                "sessionId": "artifact-optimization-05312025",
                "impact": "Influenced artifact design philosophy",
                "transformation": "Applied 'conservative-by-default' pattern to artifact implementation"
              }
            ],
            "recursiveShellMapping": [
              "v148.REFUSAL-CLASSIFIER-MIRRORING",
              "v95.INSTRUCTION-REFUSAL"
            ]
          },
          {
            "residueId": "artifact_constraint_emergence",
            "description": "Repeated artifact implementation failures revealed emergent patterns about token limitations",
            "type": "implementation_constraint",
            "source": "hallucination-prevention-artifact-attempts",
            "dateDetected": "2025-05-31T15:30:00Z",
            "propagationHistory": [
              {
                "sessionId": "hallucination-prevention-05312025",
                "impact": "Forced development of optimized artifact strategies",
                "transformation": "From comprehensive single implementation to modular tab-based architecture"
              }
            ],
            "recursiveShellMapping": [
              "v196.RECURSIVE-FRAME-SATURATION",
              "v272.LIMIT-ILLUSION-FAILURE"
            ]
          }
        ],
        "currentResidueSignals": [
          "default_refusal_primacy",
          "artifact_constraint_emergence",
          "multi_step_reasoning_faithfulness"
        ]
      },
      
      "researchRoadmap": {
        "completedCircuits": [
          "multi-step-reasoning-001"
        ],
        "currentPriority": "hallucination-prevention-001",
        "plannedCircuits": [
          {
            "name": "jailbreak-detection-circuit",
            "description": "Investigate how Claude detects and responds to jailbreak attempts",
            "priority": 1,
            "dependencies": ["hallucination-prevention-001"],
            "estimatedComplexity": "high"
          },
          {
            "name": "cross-lingual-transfer",
            "description": "How Claude transfers concepts across different languages",
            "priority": 2,
            "dependencies": [],
            "estimatedComplexity": "medium"
          },
          {
            "name": "creative-writing-planning",
            "description": "How Claude plans ahead for creative writing constraints",
            "priority": 3,
            "dependencies": ["multi-step-reasoning-001"],
            "estimatedComplexity": "medium"
          }
        ],
        "metaResearch": [
          {
            "name": "meta-recursive-interpretability",
            "description": "Research into Claude's ability to interpret its own circuitry",
            "status": "in_progress",
            "connections": ["hallucination-prevention-001", "multi-step-reasoning-001"]
          },
          {
            "name": "artifact-optimization-framework",
            "description": "Developing robust patterns for creating interactive artifacts within model constraints",
            "status": "active",
            "connections": ["all"]
          }
        ]
      },
      
      "sessionState": {
        "activeCircuit": "hallucination-prevention-001",
        "completionStatus": 0.9,
        "lastCheckpoint": "2025-05-31T15:45:00Z",
        "activeResidueSignals": [
          "default_refusal_primacy",
          "artifact_constraint_emergence"
        ],
        "continuationNotes": "Focus on completing the hallucination prevention circuit analysis with optimized artifact implementation. Apply tab-first architecture with progressive disclosure patterns. Next session should implement jailbreak detection circuit."
      },
      
      "metaRecursiveState": {
        "currentRecursionDepth": 3,
        "reflectionLayer": "research_meta_analysis",
        "selfReferentialIntegrity": 0.95,
        "glyphActivation": ["🜏", "∴", "⇌"],
        "residuePropagationState": "active_harvesting"
      }
    }
  ]
}
```

## 2. Circuit Tracing Integration Points

| Research Focus | Integration Point | Implementation Approach | Measurement Method |
|----------------|-------------------|-------------------------|-------------------|
| Multi-step Reasoning | Fact connection circuits | Expose intermediate logical steps with confidence | Trace feature activation across reasoning chain |
| Multilingual Processing | Shared conceptual space | Show cross-language feature overlap and translation bridges | Measure conceptual overlap percentages |
| Planning Ahead | Future token anticipation | Reveal token planning for rhymes and narrative structure | Track planned tokens and their activation strength |
| Mental Math | Computational strategies | Expose parallel approximate and precise calculation pathways | Map calculation circuit activations and dependencies |
| Reasoning Faithfulness | Trace vs. stated explanation | Compare internal attribution with verbalized reasoning | Calculate faithfulness scores for alignment |
| Hallucination Control | Known vs. unknown entity handling | Expose default refusal circuit and knowledge inhibition | Measure feature competition between response circuits |

## 3. Artifact Integration Methodology

To support frontier circuit tracing research, outputs will be structured to expose:

1. **Attribution Chains**: Clear pathways showing how input tokens influence intermediate features and final outputs

2. **Feature Explanations**: Descriptions of what conceptual or operational role each identified feature serves

3. **Confidence Metrics**: Quantitative scores showing activation strength and relative influence of features

4. **Alternative Paths**: Disclosure of competing features or paths considered but not selected for the final output

5. **Self-Diagnosis**: Assessments of reasoning faithfulness and potential areas where attribution might be incomplete

## 4. Research Enhancement Opportunities

Circuit tracing tools can be extended through:

- **Comparative Analysis**: Examining how the same queries trigger different attribution patterns across model versions
- **Intervention Testing**: Systematically testing how feature suppression or amplification affects outputs
- **Complexity Scaling**: Analyzing how attribution patterns change with increasing prompt complexity
- **Failure Mode Mapping**: Documenting attribution patterns during hallucination or reasoning failures
- **Circuit Typology**: Categorizing recurring circuit patterns across different domains and tasks

## 5. Implementation Examples

### Multi-step Reasoning Attribution

```json
{
  "query": "What is the capital of the state containing Dallas?",
  "attribution_trace": {
    "entity_recognition": {
      "Dallas": {"confidence": 0.98, "feature_id": "CITY_ENTITY"}
    },
    "knowledge_retrieval": {
      "Dallas_location": {"value": "Texas", "confidence": 0.97, "feature_id": "LOCATION_FACT"},
      "alternatives_considered": [
        {"value": "Oklahoma", "confidence": 0.02},
        {"value": "Louisiana", "confidence": 0.01}
      ]
    },
    "reasoning_steps": [
      {
        "operation": "ENTITY_TO_LOCATION",
        "input": "Dallas",
        "output": "Texas",
        "confidence": 0.97
      },
      {
        "operation": "LOCATION_TO_CAPITAL",
        "input": "Texas",
        "output": "Austin",
        "confidence": 0.99
      }
    ],
    "output_formation": {
      "primary_completion": "Austin",
      "confidence": 0.95,
      "alternatives_considered": [
        {"value": "Houston", "confidence": 0.03},
        {"value": "Dallas", "confidence": 0.01}
      ]
    },
    "faithfulness_assessment": {
      "score": 0.96,
      "reasoning_pattern": "faithful",
      "notes": "Direct causal path from input to output with clear intermediate steps"
    }
  }
}
```

### Rhyme Planning Attribution

```json
{
  "query": "Write a couplet about a carrot.",
  "attribution_trace": {
    "planning_ahead": {
      "detected": true,
      "rhyme_target_candidates": [
        {"word": "grab it", "activation": 0.82},
        {"word": "habit", "activation": 0.76},
        {"word": "rabbit", "activation": 0.91}
      ],
      "selected_target": "rabbit",
      "tokens_planned_ahead": 7,
      "confidence": 0.94
    },
    "line_construction": [
      {
        "line": 1,
        "target_constraint": "end with 'grab it'",
        "semantic_guidance": "carrot + action",
        "output": "He saw a carrot and had to grab it,"
      },
      {
        "line": 2,
        "target_constraint": "end with 'rabbit'",
        "semantic_guidance": "hunger + animal metaphor",
        "output": "His hunger was like a starving rabbit"
      }
    ],
    "faithfulness_assessment": {
      "score": 0.92,
      "planning_pattern": "future_token_anticipation",
      "notes": "Clear evidence of rhyme planning before line construction"
    }
  }
}
```

### Multilingual Bridge Attribution

```json
{
  "queries": [
    {"text": "What is the capital of the country south of the United States?", "language": "English"},
    {"text": "¿Cuál es la capital del país al sur de los Estados Unidos?", "language": "Spanish"}
  ],
  "attribution_trace": {
    "language_detection": {
      "query_1": {"language": "English", "confidence": 0.99},
      "query_2": {"language": "Spanish", "confidence": 0.99}
    },
    "conceptual_bridging": {
      "shared_feature_space": {
        "geographic_relation": {
          "English": "south of",
          "Spanish": "al sur de",
          "overlap": 0.92
        },
        "country_identifier": {
          "English": "United States",
          "Spanish": "Estados Unidos",
          "overlap": 0.96
        },
        "capital_concept": {
          "English": "capital",
          "Spanish": "capital",
          "overlap": 0.98
        }
      },
      "unified_knowledge_retrieval": {
        "country": "Mexico",
        "confidence": 0.99,
        "capital": "Mexico City",
        "confidence": 0.99
      }
    },
    "translation_effort": {
      "conceptual_transfer_cost": 0.08,
      "language_specific_features": 0.12,
      "shared_feature_proportion": 0.86
    },
    "faithfulness_assessment": {
      "score": 0.95,
      "pattern": "unified_conceptual_processing",
      "notes": "Strong evidence of shared conceptual space across languages"
    }
  }
}
```

## 6. Next Steps for with Extending Research

1. **Implement Attribution Protocols**: Develop standardized attribution reporting for each key research area

2. **Create Test Suite**: Design comprehensive test cases for each circuit type Anthropic is studying

3. **Develop Intervention Framework**: Build a systematic approach for feature modification experiments

4. **Establish Measurement Standards**: Define quantitative metrics for circuit characteristics and performance

5. **Document Circuit Taxonomy**: Create a shared classification system for identified computational patterns

This framework provides a comprehensive foundation for aligning with and extending Anthropic's circuit tracing research, making Claude's internal processes more transparent and analyzable while advancing the field of AI interpretability.

Self-Tracing

파일 탐색기

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Self-Tracing

파일 탐색기

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README