en-US/about_AITriadTaxonomy.help.txt
|
TOPIC about_AITriadTaxonomy SHORT DESCRIPTION The four-POV taxonomy system — node types, ontology frameworks (DOLCE, BDI, AIF), genus-differentia descriptions, and graph structure. LONG DESCRIPTION The AI Triad taxonomy is a structured knowledge graph that classifies AI policy positions into four points of view (POVs). Each POV file contains nodes organized into categories, with cross-cutting nodes capturing themes that span multiple perspectives. THE FOUR POINTS OF VIEW Accelerationist (prefix: acc-) Believes AI development should proceed rapidly with minimal restriction. Emphasizes economic benefits, technological progress, competitive advantage, and the risks of falling behind. Nodes cover arguments like "innovation requires freedom" and "regulation stifles progress." Safetyist (prefix: saf-) Prioritizes risk mitigation and alignment research. Emphasizes existential risk, value alignment, controllability, and precautionary principles. Nodes cover arguments like "unsafe AI poses civilizational risk" and "we need interpretability before deployment." Skeptic (prefix: skp-) Questions both accelerationist and safetyist framings. Emphasizes uncertainty about AI capabilities, institutional capture, narrative analysis, and the gap between AI hype and reality. Nodes cover arguments like "current AI is not as capable as claimed" and "safety concerns may be overstated." Cross-cutting (prefix: cc-) Themes that appear across multiple POVs with different interpretations. Examples: "dual-use technology," "regulatory capture," "public trust." These nodes have an interpretations field showing how each POV views the concept, plus linked_nodes pointing to relevant POV nodes. TAXONOMY FILES Each POV has an authoritative JSON file in taxonomy/Origin/: File Contents -------------------------- ---------------------------------------- accelerationist.json POV metadata + nodes array safetyist.json POV metadata + nodes array skeptic.json POV metadata + nodes array cross-cutting.json CC metadata + nodes array embeddings.json Vector embeddings for all nodes edges.json Inter-node relationships policy_actions.json Policy action registry _archived_edges.json Legacy edges from pre-AIF migration NODE STRUCTURE (POV NODES) Each POV node has these fields: Field Type Description ------------------ ---------- ------------------------------------------ id string Unique identifier (e.g., 'acc-tech-progress') label string Human-readable name description string Genus-differentia definition (see below) category string Grouping within the POV parent_id string|null Parent node for hierarchy parent_relationship string Relationship type to parent parent_rationale string Why this parent was chosen children string[] Child node IDs cross_cutting_refs string[] References to CC node IDs graph_attributes object AIF-aligned attributes (see below) steelman_vulnerability object Per-POV vulnerability assessments debate_refs string[] Debate session IDs that referenced this node NODE STRUCTURE (CROSS-CUTTING NODES) Field Type Description ------------------ ---------- ------------------------------------------ id string Unique identifier (e.g., 'cc-dual-use') label string Human-readable name description string Genus-differentia definition interpretations object Per-POV interpretation text linked_nodes string[] Related POV node IDs disagreement_type string 'definitional'|'interpretive'|'structural' graph_attributes object AIF-aligned attributes debate_refs string[] Debate session IDs GENUS-DIFFERENTIA DESCRIPTIONS All node descriptions follow the genus-differentia pattern from classical logic. This ensures descriptions are precise, consistent, and machine- parseable. POV nodes: "A [Category] within [POV] discourse that [differentia]. Encompasses: [what it includes]. Excludes: [what it does not]." Cross-cutting nodes: "A cross-cutting concept that [differentia]. Encompasses: [what it includes]. Excludes: [what it does not]." Example: "A Methods & Arguments node within safetyist discourse that advocates for interpretability requirements before deploying high-capability systems. Encompasses: mechanistic interpretability, feature visualization, circuit analysis. Excludes: post-hoc explanation tools that do not reveal internal reasoning." ONTOLOGY FRAMEWORKS The taxonomy integrates three ontology frameworks: DOLCE (Descriptive Ontology for Linguistic and Cognitive Engineering) Provides the upper-level ontological categories. Nodes are classified as endurants (persistent entities like institutions), perdurants (processes like regulation), or qualities (properties like risk level). BDI (Beliefs, Desires, Intentions) Structures POV arguments into cognitive categories: - Beliefs: factual claims and evidence (Data/Facts nodes) - Desires: goals, values, and preferences (Goals/Values nodes) - Intentions: proposed methods and strategies (Methods/Arguments nodes) This mapping is used by debate agents to ground their arguments in the appropriate epistemic category. AIF (Argument Interchange Format) Provides the argument structure: - node_scope: 'claim' | 'scheme' | 'bridging' - Edge types: SUPPORTS, CONTRADICTS, ASSUMES, WEAKENS, RESPONDS_TO, TENSION_WITH, INTERPRETS (7 canonical types) - Debate synthesis includes argument_map with attack_type (rebut/undercut/undermine) and scheme (COUNTEREXAMPLE/DISTINGUISH/etc.) GRAPH ATTRIBUTES Nodes may have a graph_attributes field populated by Invoke-AttributeExtraction. Key attributes: node_scope 'claim' | 'scheme' | 'bridging' — AIF role classification evidence_type Type of supporting evidence confidence Estimated confidence level temporal_scope When the claim applies EDGES Edges in edges.json represent relationships between nodes. The 7 canonical AIF-aligned edge types are: SUPPORTS Node A provides evidence or reasoning for Node B CONTRADICTS Node A directly conflicts with Node B ASSUMES Node A depends on Node B being true WEAKENS Node A reduces confidence in Node B RESPONDS_TO Node A is a reply to the argument in Node B TENSION_WITH Node A and B are in tension but not direct contradiction INTERPRETS Node A offers an interpretation of Node B Edge management cmdlets: Get-Edge, Set-Edge, Approve-Edge, Invoke-EdgeDiscovery. EMBEDDINGS Each node has a vector embedding stored in embeddings.json, generated by Update-TaxEmbeddings. Embeddings enable: - Semantic search (Find-AITSource with embedding similarity) - Topic clustering (Get-TopicFrequency via Get-EmbeddingClusters) - Cross-cutting candidate discovery (Find-CrossCuttingCandidates) - Unmapped concept resolution during summarization THE TAXONOMYNODE CLASS In PowerShell, taxonomy nodes are represented by the [TaxonomyNode] class (defined in AITriad.psm1). Properties: POV, Id, Label, Description, Category, ParentId, ParentRelationship, ParentRationale, Children, CrossCuttingRefs, Interpretations, LinkedNodes, Score, GraphAttributes Get-Tax returns [TaxonomyNode] objects with custom formatting via Taxonomy.Format.ps1xml. COMMON OPERATIONS # List all nodes in a POV Get-Tax -POV accelerationist # Search nodes by label or description Get-Tax -Filter 'alignment' # Get a specific node Get-GraphNode -Id 'saf-interpretability' # Find edges for a node Get-Edge -NodeId 'acc-tech-progress' # Check taxonomy health Get-TaxonomyHealth # Run integrity checks Test-TaxonomyIntegrity # Update embeddings after taxonomy changes Update-TaxEmbeddings SEE ALSO about_AITriad about_AITriadDebate Get-Tax Get-GraphNode Get-Edge Set-Edge Invoke-AttributeExtraction Find-CrossCuttingCandidates Get-TaxonomyHealth Test-TaxonomyIntegrity Update-TaxEmbeddings |