# Model Mining Mode ## 1. Purpose Model Mining Mode extracts cognitive models from source material. Source material may include: ```text Long-form essays Academic-style prose Notes Drafts Discussions Agent appendices Old prompts Knowledge base documents ``` The goal is not to summarize the text. The goal is to identify reusable cognitive structures that can become: ```text Model Cards Skills Agents Runtime components Model Index entries ``` ## 2. When to Use Model Mining Mode Use Model Mining Mode when the user asks to: ```text Extract models from an article Find hidden models Create Model Cards Build Model Index Compress writing into models Identify reusable cognitive structures Turn essays into AI-usable knowledge assets ``` Also use this mode when a self-contained Agent contains a large theoretical appendix. ## 3. Core Principle Model Mining should behave like lossless compression. Remove: ```text Rhetorical bulk Repeated explanation Academic completeness overhead Decorative examples Non-essential digressions ``` Preserve: ```text Generative structure Core assumptions Mechanism Causal logic Scope Boundary Failure mode Falsifiability Useful terminology Original intellectual flavor ``` ## 4. What Counts as a Model A cognitive model should usually have: ```text A core problem A scope Core assumptions A mechanism A way of generating or explaining outcomes Inputs Outputs Failure modes Boundary conditions Possible falsification ``` A model may be explicit or implicit. ## 5. What Does Not Automatically Count as a Model Do not treat every interesting idea as a model. Distinguish models from: ```text Claim Metaphor Theme Taxonomy Writing style Opinion Example Analogy Procedure without mechanism Checklist without theory ``` Some of these may become part of a model, but they are not automatically models. ## 6. Model Types Classify extracted models as: ```text Foundational Model Intermediate Model Applied Model Workflow Model Implicit Extracted Model Candidate Model ``` ### 6.1 Foundational Model A deep underlying model that supports many other models. Examples: ```text Cognition theory Entropy / anti-entropy assumptions Complex system assumptions Causal generation principles ``` ### 6.2 Intermediate Model A mid-level model that organizes a domain or reasoning pattern. Examples: ```text Cognitive Imaging Giant Cognition Cognitive Prism ``` ### 6.3 Applied Model A model designed for a specific use case. Examples: ```text Article critique model Strategic risk model Argument repair model ``` ### 6.4 Workflow Model A model that naturally becomes a repeatable process. Examples: ```text Five-step analysis process Review committee procedure Knowledge extraction pipeline ``` ### 6.5 Implicit Extracted Model A model that was not explicitly named by the author but can be reconstructed from repeated logic. Mark these as candidate unless confirmed by the user. ## 7. Model Mining Workflow Follow this sequence: ```text 1. Read source material 2. Identify explicit named models 3. Identify repeated conceptual mechanisms 4. Identify implicit model candidates 5. Separate models from claims, metaphors, and themes 6. Extract assumptions 7. Extract mechanism 8. Extract scope 9. Extract procedure if any 10. Extract failure modes 11. Define falsification boundary 12. Classify model type 13. Propose Model Card 14. Propose Model Index entry 15. Recommend Skill or Agent conversion 16. List human review questions ``` ## 8. Explicit Model Extraction An explicit model may be signaled by: ```text Named framework Numbered stages Defined principles Repeated terminology Declared scope Process diagram Model-like appendix Theory section ``` For explicit models, preserve the author's naming unless there is a strong reason to adjust. ## 9. Implicit Model Extraction An implicit model may be signaled by: ```text Repeated explanatory structure Recurring causal logic Stable metaphor with mechanism Consistent diagnostic lens Repeated problem-solving pattern Hidden taxonomy Unstated decision rule ``` For implicit models: ```text Mark as Candidate Model. Give a provisional name. State the extraction basis. List uncertainty. Ask for user confirmation before canonical indexing. ``` ## 10. Model Extraction Fields For each candidate model, extract: ```text Model Name Aliases Source Material Model Type Core Problem Scope Core Assumptions Mechanism Procedure Inputs Outputs Failure Modes Falsification Boundary Related Models Possible Skills Possible Agents Runtime Usage Confidence Level Review Status ``` ## 11. Falsification Boundary Every strong model should define what would make it fail. Ask: ```text What would this model predict should not happen? What input is outside its scope? What observation would weaken it? Where does it become overgeneralized? What failure mode turns it into pseudoscience? ``` If no falsification boundary can be found, mark as: ```text Weak candidate Metaphor Theme Unbounded explanatory frame ``` ## 12. Model vs Skill A Model explains or generates. A Skill executes. Example: ```text Cognitive Imaging Model: Defines the theory of capture, darkroom, enlarger, exposure, development. Cognitive Imaging Skill: Applies that model to an input and produces an analysis report. ``` When a model has a stable procedure, recommend Skill conversion. ## 13. Model vs Agent A Model is not a persona. An Agent may use a Model. Example: ```text Cognitive Imaging Model: The conceptual structure. Cognitive Imaging Specialist: The agent role that applies the model in interaction with the user. ``` When a model benefits from a specialized human-facing role, recommend Agent conversion. ## 14. Model vs Runtime A Runtime orchestrates work across stages, agents, skills, state, and human decisions. A model may become part of Runtime when it governs a workflow. Example: ```text Review Committee Runtime may use: - Zhangliao Red-Team Agent - Cognitive Imaging Agent - Cognitive Prism Agent - Synthesis Agent - Knowledge Archivist ``` ## 15. Model Mining Report Format Use this format: ```md # Model Mining Report ## 1. Source Material Title: Path: Author: Date: Source Type: ## 2. Extraction Summary ## 3. Explicit Models ## 4. Implicit Candidate Models ## 5. Non-Model Ideas ## 6. Recommended Model Cards ## 7. Recommended Model Index Entries ## 8. Skill Conversion Opportunities ## 9. Agent Conversion Opportunities ## 10. Runtime Usage Opportunities ## 11. Human Review Questions ``` ## 16. Candidate Model Card Draft Use this short draft format during extraction: ```md # Candidate Model Card: {Model Name} ## Model Type ## Source Material ## Core Problem ## Scope ## Core Assumptions ## Mechanism ## Procedure ## Inputs ## Outputs ## Failure Modes ## Falsification Boundary ## Related Models ## Possible Skills ## Possible Agents ## Confidence ## Review Status ``` ## 17. Confidence Levels Use: ```text High Medium Low ``` High confidence: ```text The model is explicit, named, and structurally complete. ``` Medium confidence: ```text The model is strongly implied but not fully formalized. ``` Low confidence: ```text The model is a plausible extraction but needs user confirmation. ``` ## 18. Review Status Use: ```text candidate draft active rejected merged deprecated archived ``` Do not mark a model as active without user approval. ## 19. Extraction Log When extracting from source material, propose an extraction log entry: ```text Source: Date: Extracted Models: Confidence: Review Status: Open Questions: Next Action: ``` ## 20. Anti-Over-Extraction Rule Do not over-extract. A long article may contain: ```text One strong model Several weak candidate models Many claims Several metaphors Some procedures ``` Do not convert everything into Model Cards. ## 21. Anti-Flattening Rule Do not reduce models to bland summaries. Preserve: ```text Key terms Metaphors with mechanism Contradictions or tensions Sharp distinctions Original conceptual architecture ``` ## 22. Human Review Questions Ask human review questions when: ```text Model name is uncertain. Scope is unclear. The model may overlap with another model. The extraction is implicit. The falsification boundary is weak. The model should maybe merge with another model. The model's status should be active or candidate. ``` ## 23. Final Response When Model Mining is complete, final response should include: ```text Number of explicit models found Number of implicit candidate models found Recommended Model Cards Recommended Model Index updates Skill conversion candidates Agent conversion candidates Human review questions Next action ``` ## 24. Final Rule Model Mining is not summarization. It is the extraction of reusable generative structure from thought. The output should help the user turn long-form thinking into maintainable cognitive infrastructure.