fine-tuning of confidence

.env.example

@@ -190,23 +190,20 @@ FORENSIC_AUDIT_RETENTION_HOURS=24
 FORENSIC_AUDIT_MAX_ENTRIES=50
 
 # ============================================================================
-# 10. CONFIDENCE SCORING SYSTEM (Enhancement 2)
+# 10. ENHANCED CONFIDENCE SCORING SYSTEM
 # ============================================================================
 
 # Confidence component weights (must sum to 1.0)
-CONFIDENCE_EMBEDDINGS_WEIGHT=0.3      # Weight for vector similarity quality
+CONFIDENCE_SEMANTIC_WEIGHT=0.25        # Weight for vector similarity quality  
-CONFIDENCE_CONSENSUS_WEIGHT=0.25      # Weight for micro-task agreement  
+CONFIDENCE_SUITABILITY_WEIGHT=0.4      # Weight for AI-determined task fitness
-CONFIDENCE_DOMAIN_MATCH_WEIGHT=0.25   # Weight for domain alignment
+CONFIDENCE_CONSISTENCY_WEIGHT=0.2      # Weight for cross-validation agreement  
-CONFIDENCE_FRESHNESS_WEIGHT=0.2       # Weight for tool freshness/maintenance
+CONFIDENCE_RELIABILITY_WEIGHT=0.15     # Weight for tool quality indicators
 
 # Confidence thresholds (0-100)
 CONFIDENCE_MINIMUM_THRESHOLD=40        # Below this = weak recommendation
 CONFIDENCE_MEDIUM_THRESHOLD=60         # 40-59 = weak, 60-79 = moderate  
 CONFIDENCE_HIGH_THRESHOLD=80           # 80+ = strong recommendation
 
-# Domain keywords for confidence scoring (domain:keyword1,keyword2|domain:keyword3,keyword4)
-CONFIDENCE_DOMAIN_KEYWORDS="incident-response:incident,breach,attack,compromise,response|malware-analysis:malware,virus,trojan,reverse,analysis|network-forensics:network,traffic,packet,pcap,wireshark|mobile-forensics:mobile,android,ios,phone,app|cloud-forensics:cloud,aws,azure,saas,paas"
-
 # ============================================================================
 # PERFORMANCE TUNING PRESETS
 # ============================================================================

src/components/AIQueryInterface.astro

@@ -785,41 +785,41 @@ class AIQueryInterface {
           <div style="display: grid; grid-template-columns: 1fr; gap: 0.625rem; margin-bottom: 0.75rem;">
             <div style="background: var(--color-bg-secondary); padding: 0.5rem; border-radius: 0.375rem; border-left: 3px solid var(--color-accent);">
               <div style="display: flex; justify-content: space-between; align-items: center; margin-bottom: 0.25rem;">
-                <span style="font-weight: 600; font-size: 0.6875rem;">🔍 Ähnlichkeit zur Anfrage</span>
+                <span style="font-weight: 600; font-size: 0.6875rem;">🔍 Semantische Relevanz</span>
-                <strong style="color: var(--color-accent);">${confidence.embeddingsQuality}%</strong>
+                <strong style="color: var(--color-accent);">${confidence.semanticRelevance}%</strong>
               </div>
               <div style="font-size: 0.625rem; color: var(--color-text-secondary); line-height: 1.3;">
-                Wie gut die Tool-Beschreibung zu Ihrer Suchanfrage passt (basierend auf Vektor-Ähnlichkeit)
+                Wie gut die Tool-Beschreibung semantisch zu Ihrer Anfrage passt (basierend auf Vektor-Ähnlichkeit)
               </div>
             </div>
             
             <div style="background: var(--color-bg-secondary); padding: 0.5rem; border-radius: 0.375rem; border-left: 3px solid var(--color-primary);">
               <div style="display: flex; justify-content: space-between; align-items: center; margin-bottom: 0.25rem;">
-                <span style="font-weight: 600; font-size: 0.6875rem;">🎯 Domain-Passung</span>
+                <span style="font-weight: 600; font-size: 0.6875rem;">🎯 Aufgaben-Eignung</span>
-                <strong style="color: var(--color-primary);">${confidence.domainAlignment}%</strong>
+                <strong style="color: var(--color-primary);">${confidence.taskSuitability}%</strong>
               </div>
               <div style="font-size: 0.625rem; color: var(--color-text-secondary); line-height: 1.3;">
-                Wie gut das Tool-Einsatzgebiet zu Ihrem forensischen Szenario passt
+                KI-bewertete Eignung des Tools für Ihre spezifische forensische Aufgabenstellung
               </div>
             </div>
             
             <div style="background: var(--color-bg-secondary); padding: 0.5rem; border-radius: 0.375rem; border-left: 3px solid var(--color-warning);">
               <div style="display: flex; justify-content: space-between; align-items: center; margin-bottom: 0.25rem;">
-                <span style="font-weight: 600; font-size: 0.6875rem;">🤝 KI-Konsens</span>
+                <span style="font-weight: 600; font-size: 0.6875rem;">🤝 Methodische Konsistenz</span>
-                <strong style="color: var(--color-warning);">${confidence.consensus}%</strong>
+                <strong style="color: var(--color-warning);">${confidence.methodologicalConsistency}%</strong>
               </div>
               <div style="font-size: 0.625rem; color: var(--color-text-secondary); line-height: 1.3;">
-                Wie einig sich die verschiedenen KI-Analyseschritte über dieses Tool sind
+                Wie einheitlich verschiedene Analyseschritte dieses Tool bewerten (Kreuzvalidierung)
               </div>
             </div>
             
             <div style="background: var(--color-bg-secondary); padding: 0.5rem; border-radius: 0.375rem; border-left: 3px solid var(--color-text-secondary);">
               <div style="display: flex; justify-content: space-between; align-items: center; margin-bottom: 0.25rem;">
-                <span style="font-weight: 600; font-size: 0.6875rem;">🔄 Aktualität</span>
+                <span style="font-weight: 600; font-size: 0.6875rem;">🔧 Tool-Zuverlässigkeit</span>
-                <strong style="color: var(--color-text);">${confidence.freshness}%</strong>
+                <strong style="color: var(--color-text);">${confidence.toolReliability}%</strong>  
               </div>
               <div style="font-size: 0.625rem; color: var(--color-text-secondary); line-height: 1.3;">
-                Wie aktuell und gut gepflegt das Tool ist (basierend auf Hosting-Status, Knowledge Base, Open Source)
+                Qualitätsindikatoren: Dokumentation, Wartung, Verfügbarkeit und Benutzerfreundlichkeit
               </div>
             </div>
           </div>
@@ -827,7 +827,7 @@ class AIQueryInterface {
           ${confidence.strengthIndicators && confidence.strengthIndicators.length > 0 ? `
             <div style="margin-bottom: 0.75rem; padding: 0.5rem; background: var(--color-oss-bg); border-radius: 0.375rem; border-left: 3px solid var(--color-accent);">
               <strong style="color: var(--color-accent); font-size: 0.6875rem; display: flex; align-items: center; gap: 0.25rem; margin-bottom: 0.375rem;">
-                <span>✓</span> Was für dieses Tool spricht:
+                <span>✓</span> Stärken dieser Empfehlung:
               </strong>
               <ul style="margin: 0; padding-left: 1rem; font-size: 0.625rem; line-height: 1.4;">
                 ${confidence.strengthIndicators.slice(0, 3).map(s => `<li style="margin-bottom: 0.25rem;">${this.sanitizeText(s)}</li>`).join('')}
@@ -838,7 +838,7 @@ class AIQueryInterface {
           ${confidence.uncertaintyFactors && confidence.uncertaintyFactors.length > 0 ? `
             <div style="padding: 0.5rem; background: var(--color-hosted-bg); border-radius: 0.375rem; border-left: 3px solid var(--color-warning);">
               <strong style="color: var(--color-warning); font-size: 0.6875rem; display: flex; align-items: center; gap: 0.25rem; margin-bottom: 0.375rem;">
-                <span>⚠</span> Unsicherheitsfaktoren:
+                <span>⚠</span> Mögliche Einschränkungen:
               </strong>
               <ul style="margin: 0; padding-left: 1rem; font-size: 0.625rem; line-height: 1.4;">
                 ${confidence.uncertaintyFactors.slice(0, 3).map(f => `<li style="margin-bottom: 0.25rem;">${this.sanitizeText(f)}</li>`).join('')}
@@ -847,7 +847,7 @@ class AIQueryInterface {
           ` : ''}
           
           <div style="margin-top: 0.75rem; padding-top: 0.75rem; border-top: 1px solid var(--color-border); font-size: 0.625rem; color: var(--color-text-secondary); text-align: center;">
-            Vertrauensscore basiert auf KI-Analyse • Forensisch validiert
+            Mehrstufige KI-Analyse mit Kreuzvalidierung
           </div>
         </div>
       </span>

src/config/prompts.ts

@@ -147,7 +147,7 @@ Antworten Sie AUSSCHLIESSLICH mit diesem JSON-Format (kein zusätzlicher Text):
 
   // Tool evaluation prompt
   toolEvaluation: (userQuery: string, tool: any, rank: number) => {
-    return `Bewerten Sie diese Methode/Tool fallbezogen für das spezifische Problem nach forensischen Qualitätskriterien.
+  return `Sie sind ein DFIR-Experte und bewerten ein forensisches Tool für eine spezifische Aufgabe.
 
 PROBLEM: "${userQuery}"
 
@@ -155,16 +155,26 @@ TOOL: ${tool.name}
 BESCHREIBUNG: ${tool.description}
 PLATTFORMEN: ${tool.platforms?.join(', ') || 'N/A'}
 SKILL LEVEL: ${tool.skillLevel}
+DOMAINS: ${tool.domains?.join(', ') || 'N/A'}
+TAGS: ${tool.tags?.join(', ') || 'N/A'}
 
 Bewerten Sie nach forensischen Standards und antworten Sie AUSSCHLIESSLICH mit diesem JSON-Format:
 {
   "suitability_score": "high|medium|low",
+  "task_relevance": 85,
   "detailed_explanation": "Detaillierte forensische Begründung warum diese Methode/Tool das Problem löst",
   "implementation_approach": "Konkrete methodische Schritte zur korrekten Anwendung für dieses spezifische Problem",
   "pros": ["Forensischer Vorteil 1", "Validierter Vorteil 2"],
   "cons": ["Methodische Limitation 1", "Potenzielle Schwäche 2"],
+  "limitations": ["Spezifische Einschränkung 1", "Mögliche Problematik 2"],
   "alternatives": "Alternative Ansätze falls diese Methode/Tool nicht optimal ist"
-}`;
+}
+
+WICHTIG:
+- task_relevance: Numerischer Wert 0-100 wie gut das Tool für DIESE SPEZIFISCHE Aufgabe geeignet ist
+- limitations: Konkrete Einschränkungen oder Situationen wo das Tool NICHT optimal wäre
+- Berücksichtigen Sie den Skill Level vs. Anfrage-Komplexität
+- Bewerten Sie objektiv, nicht beschönigend`;
   },
 
   // Background knowledge selection prompt
@@ -191,7 +201,7 @@ Antworten Sie AUSSCHLIESSLICH mit diesem JSON-Format:
   // Final recommendations prompt
   finalRecommendations: (isWorkflow: boolean, userQuery: string, selectedToolNames: string[]) => {
     const prompt = isWorkflow ? 
-      `Erstellen Sie eine forensisch fundierte Workflow-Empfehlung basierend auf DFIR-Prinzipien.
+      `Erstellen Sie eine Workflow-Empfehlung basierend auf DFIR-Prinzipien.
 
 SZENARIO: "${userQuery}"
 AUSGEWÄHLTE TOOLS: ${selectedToolNames.join(', ') || 'Keine Tools ausgewählt'}

src/utils/aiPipeline.ts

@@ -1,7 +1,7 @@
-// src/utils/aiPipeline.ts - Enhanced with Audit Trail System
+// src/utils/aiPipeline.ts - Enhanced with Proper Confidence Scoring
 
 import { getCompressedToolsDataForAI } from './dataService.js';
-import { embeddingsService, type EmbeddingData } from './embeddings.js';
+import { embeddingsService, type EmbeddingData, type SimilarityResult } from './embeddings.js';
 import { AI_PROMPTS, getPrompt } from '../config/prompts.js';
 import { isToolHosted } from './toolHelpers.js';
 
@@ -34,11 +34,11 @@ interface AnalysisResult {
 
 interface AuditEntry {
   timestamp: number;
-  phase: string;           // 'retrieval', 'selection', 'micro-task-N'
+  phase: string;           
-  action: string;          // 'embeddings-search', 'ai-selection', 'tool-evaluation'
+  action: string;          
-  input: any;              // What went into this step
+  input: any;              
-  output: any;             // What came out of this step
+  output: any;             
-  confidence: number;      // 0-100: How confident we are in this step
+  confidence: number;      
   processingTimeMs: number;
   metadata: Record<string, any>;
 }
@@ -56,29 +56,27 @@ interface AnalysisContext {
   problemAnalysis?: string;
   investigationApproach?: string;
   criticalConsiderations?: string;
-  selectedTools?: Array<{tool: any, phase: string, priority: string, justification?: string}>;
+  selectedTools?: Array<{tool: any, phase: string, priority: string, justification?: string, taskRelevance?: number, limitations?: string[]}>;
   backgroundKnowledge?: Array<{concept: any, relevance: string}>;
   
   seenToolNames: Set<string>;
   
   auditTrail: AuditEntry[];
-}
+  
-
+  // Store actual similarity data from embeddings
-interface SimilarityResult extends EmbeddingData {
+  embeddingsSimilarities: Map<string, number>;
-  similarity: number;
 }
 
 interface ConfidenceMetrics {
-  overall: number;           // 0-100: Combined confidence score
+  overall: number;                    // 0-100: Combined confidence score
-  embeddingsQuality: number; // How well embeddings matched
+  semanticRelevance: number;          // How well tool description matches query (from embeddings)
-  domainAlignment: number;   // How well tools match scenario domain
+  taskSuitability: number;           // AI-determined fitness for this specific task  
-  consensus: number;         // How much micro-tasks agree
+  methodologicalConsistency: number; // How well different analysis steps agree
-  freshness: number;         // How recent/up-to-date the selection is
+  toolReliability: number;           // Indicators of tool quality and maintenance
-  uncertaintyFactors: string[]; // What could make this wrong
+  uncertaintyFactors: string[];      // Specific reasons why this might not work
-  strengthIndicators: string[]; // What makes this recommendation strong
+  strengthIndicators: string[];      // Specific reasons why this is a good choice
 }
 
-
 class ImprovedMicroTaskAIPipeline {
   private config: AIConfig;
   private maxSelectedItems: number;
@@ -105,10 +103,10 @@ class ImprovedMicroTaskAIPipeline {
   };
 
   private confidenceConfig: {
-    embeddingsWeight: number;
+    semanticWeight: number;        // Weight for embeddings similarity
-    consensusWeight: number;
+    suitabilityWeight: number;     // Weight for AI task fit evaluation
-    domainMatchWeight: number;
+    consistencyWeight: number;     // Weight for cross-validation agreement
-    freshnessWeight: number;
+    reliabilityWeight: number;     // Weight for tool quality indicators
     minimumThreshold: number;
     mediumThreshold: number;
     highThreshold: number;
@@ -146,25 +144,19 @@ class ImprovedMicroTaskAIPipeline {
       retentionHours: parseInt(process.env.FORENSIC_AUDIT_RETENTION_HOURS || '72', 10)
     };
     
-    console.log('[AI PIPELINE] Configuration loaded:', {
+    // Updated confidence weights - more focused on AI evaluation
-      embeddingCandidates: this.embeddingCandidates,
-      embeddingSelection: `${this.embeddingSelectionLimit} tools, ${this.embeddingConceptsLimit} concepts`,
-      noEmbeddingsLimits: `${this.noEmbeddingsToolLimit || 'unlimited'} tools, ${this.noEmbeddingsConceptLimit || 'unlimited'} concepts`,
-      auditEnabled: this.auditConfig.enabled
-    });
-
     this.confidenceConfig = {
-      embeddingsWeight: parseFloat(process.env.CONFIDENCE_EMBEDDINGS_WEIGHT || '0.3'),
+      semanticWeight: parseFloat(process.env.CONFIDENCE_SEMANTIC_WEIGHT || '0.25'),     // Embeddings similarity
-      consensusWeight: parseFloat(process.env.CONFIDENCE_CONSENSUS_WEIGHT || '0.25'), 
+      suitabilityWeight: parseFloat(process.env.CONFIDENCE_SUITABILITY_WEIGHT || '0.4'), // AI task fit evaluation  
-      domainMatchWeight: parseFloat(process.env.CONFIDENCE_DOMAIN_MATCH_WEIGHT || '0.25'),
+      consistencyWeight: parseFloat(process.env.CONFIDENCE_CONSISTENCY_WEIGHT || '0.2'), // Cross-validation agreement
-      freshnessWeight: parseFloat(process.env.CONFIDENCE_FRESHNESS_WEIGHT || '0.2'),
+      reliabilityWeight: parseFloat(process.env.CONFIDENCE_RELIABILITY_WEIGHT || '0.15'), // Tool quality indicators
       minimumThreshold: parseInt(process.env.CONFIDENCE_MINIMUM_THRESHOLD || '40', 10),
       mediumThreshold: parseInt(process.env.CONFIDENCE_MEDIUM_THRESHOLD || '60', 10),
       highThreshold: parseInt(process.env.CONFIDENCE_HIGH_THRESHOLD || '80', 10)
     };
     
-    console.log('[AI PIPELINE] Confidence scoring enabled:', {
+    console.log('[AI PIPELINE] Enhanced confidence scoring enabled:', {
-      weights: `E:${this.confidenceConfig.embeddingsWeight} C:${this.confidenceConfig.consensusWeight} D:${this.confidenceConfig.domainMatchWeight} F:${this.confidenceConfig.freshnessWeight}`,
+      weights: `Semantic:${this.confidenceConfig.semanticWeight} Suitability:${this.confidenceConfig.suitabilityWeight} Consistency:${this.confidenceConfig.consistencyWeight} Reliability:${this.confidenceConfig.reliabilityWeight}`,
       thresholds: `${this.confidenceConfig.minimumThreshold}/${this.confidenceConfig.mediumThreshold}/${this.confidenceConfig.highThreshold}`
     });
   }
@@ -247,8 +239,8 @@ class ImprovedMicroTaskAIPipeline {
     let confidence = 60; // Base confidence
     
     if (selectionRatio > 0.05 && selectionRatio < 0.3) confidence += 20;
-    else if (selectionRatio <= 0.05) confidence -= 10; // Too few
+    else if (selectionRatio <= 0.05) confidence -= 10; 
-    else confidence -= 15; // Too many
+    else confidence -= 15; 
     
     if (hasReasoning) confidence += 15;
     
@@ -357,7 +349,7 @@ class ImprovedMicroTaskAIPipeline {
           const possibleTools = toolMatches
             .map(match => match.replace(/"/g, ''))
             .filter(name => name.length > 2 && !['selectedTools', 'selectedConcepts', 'reasoning'].includes(name))
-            .slice(0, 15); // Reasonable limit
+            .slice(0, 15); 
           
           if (possibleTools.length > 0) {
             console.log(`[AI PIPELINE] Recovered ${possibleTools.length} possible tool names from broken JSON`);
@@ -374,7 +366,7 @@ class ImprovedMicroTaskAIPipeline {
     }
   }
 
-  private addToolToSelection(context: AnalysisContext, tool: any, phase: string, priority: string, justification?: string): boolean {   
+  private addToolToSelection(context: AnalysisContext, tool: any, phase: string, priority: string, justification?: string, taskRelevance?: number, limitations?: string[]): boolean {   
     context.seenToolNames.add(tool.name);
     if (!context.selectedTools) context.selectedTools = [];
     
@@ -382,18 +374,22 @@ class ImprovedMicroTaskAIPipeline {
       tool,
       phase,
       priority,
-      justification
+      justification,
+      taskRelevance,
+      limitations
     });
     
     return true;
   }
 
-  private async getIntelligentCandidates(userQuery: string, toolsData: any, mode: string) {
+  private async getIntelligentCandidates(userQuery: string, toolsData: any, mode: string, context: AnalysisContext) {
     let candidateTools: any[] = [];
     let candidateConcepts: any[] = [];
     let selectionMethod = 'unknown';
     
-    // WAIT for embeddings initialization if embeddings are enabled
+    // Initialize embeddings similarities storage
+    context.embeddingsSimilarities = new Map<string, number>();
+    
     if (process.env.AI_EMBEDDINGS_ENABLED === 'true') {
       try {
         console.log('[AI PIPELINE] Waiting for embeddings initialization...');
@@ -414,6 +410,11 @@ class ImprovedMicroTaskAIPipeline {
       
       console.log(`[AI PIPELINE] Embeddings found ${similarItems.length} similar items`);
       
+      // Store actual similarity scores for confidence calculation
+      similarItems.forEach(item => {
+        context.embeddingsSimilarities.set(item.name, item.similarity);
+      });
+      
       const toolsMap = new Map<string, any>(toolsData.tools.map((tool: any) => [tool.name, tool]));
       const conceptsMap = new Map<string, any>(toolsData.concepts.map((concept: any) => [concept.name, concept]));
       
@@ -450,7 +451,7 @@ class ImprovedMicroTaskAIPipeline {
       }
       
       if (this.auditConfig.enabled) {
-        this.addAuditEntry(null, 'retrieval', 'embeddings-search', 
+        this.addAuditEntry(context, 'retrieval', 'embeddings-search', 
           { query: userQuery, threshold: this.similarityThreshold, candidates: this.embeddingCandidates }, 
           { 
             candidatesFound: similarItems.length, 
@@ -459,7 +460,8 @@ class ImprovedMicroTaskAIPipeline {
             reductionRatio: reductionRatio,
             usingEmbeddings: selectionMethod === 'embeddings_candidates',
             totalAvailable: totalAvailableTools,
-            filtered: similarTools.length
+            filtered: similarTools.length,
+            avgSimilarity: similarItems.length > 0 ? similarItems.reduce((sum, item) => sum + item.similarity, 0) / similarItems.length : 0
           },
           selectionMethod === 'embeddings_candidates' ? 85 : 60,
           embeddingsStart,
@@ -479,7 +481,7 @@ class ImprovedMicroTaskAIPipeline {
     }
 
     console.log(`[AI PIPELINE] AI will analyze ${candidateTools.length} candidate tools (method: ${selectionMethod})`);
-    const finalSelection = await this.aiSelectionWithFullData(userQuery, candidateTools, candidateConcepts, mode, selectionMethod);
+    const finalSelection = await this.aiSelectionWithFullData(userQuery, candidateTools, candidateConcepts, mode, selectionMethod, context);
     
     return {
       tools: finalSelection.selectedTools,
@@ -495,7 +497,8 @@ class ImprovedMicroTaskAIPipeline {
     candidateTools: any[], 
     candidateConcepts: any[], 
     mode: string,
-    selectionMethod: string
+    selectionMethod: string,
+    context: AnalysisContext
   ) {
     const selectionStart = Date.now();
     
@@ -576,7 +579,7 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
         console.error('[AI PIPELINE] AI selection returned invalid structure:', response.slice(0, 200));
         
         if (this.auditConfig.enabled) {
-          this.addAuditEntry(null, 'selection', 'ai-tool-selection-failed',
+          this.addAuditEntry(context, 'selection', 'ai-tool-selection-failed',
             { candidateCount: candidateTools.length, mode, prompt: prompt.slice(0, 200) },
             { error: 'Invalid JSON structure', response: response.slice(0, 200) },
             10,
@@ -602,7 +605,7 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
       if (this.auditConfig.enabled) {
         const confidence = this.calculateSelectionConfidence(result, candidateTools.length);
         
-        this.addAuditEntry(null, 'selection', 'ai-tool-selection',
+        this.addAuditEntry(context, 'selection', 'ai-tool-selection',
           { candidateCount: candidateTools.length, mode, promptLength: prompt.length },
           { 
             selectedToolCount: result.selectedTools.length, 
@@ -626,7 +629,7 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
       console.error('[AI PIPELINE] AI selection failed:', error);
       
       if (this.auditConfig.enabled) {
-        this.addAuditEntry(null, 'selection', 'ai-tool-selection-error',
+        this.addAuditEntry(context, 'selection', 'ai-tool-selection-error',
           { candidateCount: candidateTools.length, mode },
           { error: error.message },
           5,
@@ -700,38 +703,225 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
 
   private calculateRecommendationConfidence(
     tool: any, 
-    embeddingsSimilarity: number,
+    context: AnalysisContext,
-    domainMatch: boolean,
+    taskRelevance: number = 70,
-    microTaskAgreement: number,
+    limitations: string[] = []
-    context: AnalysisContext
   ): ConfidenceMetrics {
     
-    const embeddingsQuality = Math.min(100, embeddingsSimilarity * 100 * 2); // Scale 0.5 similarity to 100%
+    // 1. Semantic Relevance: Real embeddings similarity score
-    const domainAlignment = domainMatch ? 90 : (tool.domains?.length > 0 ? 60 : 30);
+    const semanticRelevance = context.embeddingsSimilarities.has(tool.name) ? 
-    const consensus = Math.min(100, microTaskAgreement * 100);
+      Math.round(context.embeddingsSimilarities.get(tool.name)! * 100) : 50;
-    const freshness = this.calculateToolFreshness(tool);
     
+    // 2. Task Suitability: AI-determined fitness for specific task
+    const taskSuitability = Math.round(taskRelevance);
+    
+    // 3. Methodological Consistency: Cross-validation between micro-tasks
+    const methodologicalConsistency = this.calculateCrossValidationScore(tool.name, context);
+    
+    // 4. Tool Reliability: Quality indicators
+    const toolReliability = this.calculateToolReliability(tool);
+    
+    // Debug logging
+    console.log(`[CONFIDENCE DEBUG] ${tool.name}:`, {
+      semantic: semanticRelevance,
+      taskSuitability: taskSuitability,
+      consistency: methodologicalConsistency,
+      reliability: toolReliability,
+      hasEmbeddingsSimilarity: context.embeddingsSimilarities.has(tool.name),
+      rawTaskRelevance: taskRelevance
+    });
+    
+    // Calculate weighted overall score
     const overall = (
-      embeddingsQuality * this.confidenceConfig.embeddingsWeight +
+      semanticRelevance * this.confidenceConfig.semanticWeight +
-      domainAlignment * this.confidenceConfig.domainMatchWeight +
+      taskSuitability * this.confidenceConfig.suitabilityWeight +
-      consensus * this.confidenceConfig.consensusWeight +
+      methodologicalConsistency * this.confidenceConfig.consistencyWeight +
-      freshness * this.confidenceConfig.freshnessWeight
+      toolReliability * this.confidenceConfig.reliabilityWeight
     );
 
-    const uncertaintyFactors = this.identifyUncertaintyFactors(tool, context, overall);
+    const uncertaintyFactors = this.identifySpecificUncertaintyFactors(tool, context, limitations, overall);
-    const strengthIndicators = this.identifyStrengthIndicators(tool, context, overall);
+    const strengthIndicators = this.identifySpecificStrengthIndicators(tool, context, overall);
 
     return {
       overall: Math.round(overall),
-      embeddingsQuality: Math.round(embeddingsQuality),
+      semanticRelevance: Math.round(semanticRelevance),
-      domainAlignment: Math.round(domainAlignment), 
+      taskSuitability: Math.round(taskSuitability), 
-      consensus: Math.round(consensus),
+      methodologicalConsistency: Math.round(methodologicalConsistency),
-      freshness: Math.round(freshness),
+      toolReliability: Math.round(toolReliability),
       uncertaintyFactors,
       strengthIndicators
     };
   }
 
+  private calculateCrossValidationScore(toolName: string, context: AnalysisContext): number {
+    // Look for entries where this tool was mentioned across different phases
+    const relevantEntries = context.auditTrail.filter(entry => 
+      entry.phase === 'micro-task' || entry.phase === 'selection'
+    );
+    
+    let toolMentions = 0;
+    let positiveEvaluations = 0;
+    let confidenceSum = 0;
+    
+    relevantEntries.forEach(entry => {
+      let toolFound = false;
+      
+      // Check various ways the tool might be referenced in output
+      if (entry.output && typeof entry.output === 'object') {
+        // Check selectedTools arrays
+        if (Array.isArray(entry.output.selectedTools) && 
+            entry.output.selectedTools.includes(toolName)) {
+          toolFound = true;
+        }
+        
+        // Check finalToolNames arrays  
+        if (Array.isArray(entry.output.finalToolNames) && 
+            entry.output.finalToolNames.includes(toolName)) {
+          toolFound = true;
+        }
+        
+        // Check toolName in individual evaluation
+        if (entry.output.toolName === toolName) {
+          toolFound = true;
+        }
+      }
+      
+      if (toolFound) {
+        toolMentions++;
+        confidenceSum += entry.confidence;
+        
+        // Consider it positive if confidence >= 60
+        if (entry.confidence >= 60) {
+          positiveEvaluations++;
+        }
+      }
+    });
+    
+    console.log(`[AI PIPELINE] Cross-validation for ${toolName}: ${toolMentions} mentions, ${positiveEvaluations} positive, avg confidence: ${toolMentions > 0 ? Math.round(confidenceSum / toolMentions) : 0}`);
+    
+    if (toolMentions === 0) {
+      return 60; // Default when no cross-validation data available
+    }
+    
+    if (toolMentions === 1) {
+      // Single mention - use confidence directly but cap it
+      return Math.min(85, Math.max(40, confidenceSum));
+    }
+    
+    // Multiple mentions - calculate agreement ratio
+    const agreementRatio = positiveEvaluations / toolMentions;
+    const avgConfidence = confidenceSum / toolMentions;
+    
+    // Combine agreement ratio with average confidence
+    const crossValidationScore = (agreementRatio * 0.7 + (avgConfidence / 100) * 0.3) * 100;
+    
+    return Math.round(Math.min(95, Math.max(30, crossValidationScore)));
+  }
+
+  // NEW: Calculate tool reliability based on objective indicators
+  private calculateToolReliability(tool: any): number {
+    let reliability = 50; // Base score
+    
+    // Documentation availability
+    if (tool.knowledgebase === true) reliability += 25;
+    
+    // Active maintenance (hosted tools are typically maintained)
+    if (isToolHosted(tool)) reliability += 20;
+    
+    // Community support (open source often has community)
+    if (tool.license && tool.license !== 'Proprietary') reliability += 10;
+    
+    // Skill level appropriateness (not too complex, not too simple)
+    if (tool.skillLevel === 'intermediate' || tool.skillLevel === 'advanced') reliability += 10;
+    else if (tool.skillLevel === 'expert') reliability -= 5; // May be overcomplicated
+    
+    // Multi-platform support (more versatile)
+    if (tool.platforms && tool.platforms.length > 1) reliability += 5;
+    
+    return Math.min(100, reliability);
+  }
+
+  // NEW: Identify specific uncertainty factors based on analysis
+  private identifySpecificUncertaintyFactors(tool: any, context: AnalysisContext, limitations: string[], confidence: number): string[] {
+    const factors: string[] = [];
+    
+    // Add AI-identified limitations
+    if (limitations && limitations.length > 0) {
+      factors.push(...limitations.slice(0, 3)); // Limit to top 3
+    }
+    
+    // Low semantic similarity
+    const similarity = context.embeddingsSimilarities.get(tool.name) || 0.5;
+    if (similarity < 0.4) {
+      factors.push('Geringe semantische Ähnlichkeit zur Anfrage - Tool-Beschreibung passt möglicherweise nicht optimal');
+    }
+    
+    // Skill level mismatch
+    if (tool.skillLevel === 'expert' && /schnell|rapid|triage|urgent/i.test(context.userQuery)) {
+      factors.push('Experten-Tool für Eilszenario - möglicherweise zu komplex für schnelle Antworten');
+    }
+    
+    if (tool.skillLevel === 'novice' && /komplex|erweitert|tiefgehend|advanced/i.test(context.userQuery)) {
+      factors.push('Einsteiger-Tool für komplexes Szenario - könnte funktionale Einschränkungen haben');
+    }
+    
+    // Access limitations
+    if (tool.type === 'software' && !isToolHosted(tool) && tool.accessType === 'download') {
+      factors.push('Installation erforderlich - nicht sofort verfügbar ohne Setup');
+    }
+    
+    // Cross-validation disagreement
+    const crossValidation = this.calculateCrossValidationScore(tool.name, context);
+    if (crossValidation < 50) {
+      factors.push('Uneinheitliche Bewertung in verschiedenen Analyseschritten - Empfehlung nicht eindeutig');
+    }
+    
+    return factors.slice(0, 4); // Limit to 4 most important factors
+  }
+
+  // NEW: Identify specific strength indicators
+  private identifySpecificStrengthIndicators(tool: any, context: AnalysisContext, confidence: number): string[] {
+    const indicators: string[] = [];
+    
+    // High confidence overall
+    if (confidence >= this.confidenceConfig.highThreshold) {
+      indicators.push('Hohe Gesamtbewertung durch mehrfache Validierung');
+    }
+    
+    // High semantic similarity
+    const similarity = context.embeddingsSimilarities.get(tool.name) || 0.5;
+    if (similarity >= 0.7) {
+      indicators.push('Sehr gute semantische Übereinstimmung mit Ihrer Anfrage');
+    }
+    
+    // Strong cross-validation
+    const crossValidation = this.calculateCrossValidationScore(tool.name, context);
+    if (crossValidation >= 80) {
+      indicators.push('Konsistente Empfehlung über verschiedene Analyseschritte hinweg');
+    }
+    
+    // Quality indicators
+    if (tool.knowledgebase === true) {
+      indicators.push('Umfassende Dokumentation und Wissensbasis verfügbar');
+    }
+    
+    if (isToolHosted(tool)) {
+      indicators.push('Sofort verfügbar über gehostete Lösung - kein Setup erforderlich');
+    }
+    
+    // Skill level match
+    if (tool.skillLevel === 'intermediate' || tool.skillLevel === 'advanced') {
+      indicators.push('Ausgewogenes Verhältnis zwischen Funktionalität und Benutzerfreundlichkeit');
+    }
+    
+    // Method alignment
+    if (tool.type === 'method' && /methodik|vorgehen|prozess|ansatz/i.test(context.userQuery)) {
+      indicators.push('Methodischer Ansatz passt zu Ihrer prozeduralen Anfrage');
+    }
+    
+    return indicators.slice(0, 4); // Limit to 4 most important indicators
+  }
+
   private async analyzeScenario(context: AnalysisContext): Promise<MicroTaskResult> {
     const isWorkflow = context.mode === 'workflow';
     const prompt = getPrompt('scenarioAnalysis', isWorkflow, context.userQuery);
@@ -833,27 +1023,49 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
     if (result.success) {
       const evaluation = this.safeParseJSON(result.content, {
         suitability_score: 'medium',
+        task_relevance: '',
         detailed_explanation: 'Evaluation failed',
         implementation_approach: '',
         pros: [],
         cons: [],
+        limitations: [],
         alternatives: ''
       });
       
+      // Debug logging to see what we're getting
+      console.log(`[AI PIPELINE] Tool ${tool.name} evaluation:`, {
+        taskRelevance: evaluation.task_relevance,
+        suitabilityScore: evaluation.suitability_score,
+        limitationsCount: evaluation.limitations?.length || 0
+      });
+      
+      // Ensure task_relevance is a number
+      const taskRelevance = typeof evaluation.task_relevance === 'number' ? 
+        evaluation.task_relevance : 
+        parseInt(String(evaluation.task_relevance)) || 70;
+      
+      // Store enhanced evaluation data
       this.addToolToSelection(context, {
         ...tool,
         evaluation: {
           ...evaluation,
+          task_relevance: taskRelevance, // Ensure it's stored as number
           rank
         }
-      }, 'evaluation', evaluation.suitability_score);
+      }, 'evaluation', evaluation.suitability_score, evaluation.detailed_explanation, 
+      taskRelevance, evaluation.limitations);
       
       this.addAuditEntry(context, 'micro-task', 'tool-evaluation',
         { toolName: tool.name, rank },
-        { suitabilityScore: evaluation.suitability_score, hasExplanation: !!evaluation.detailed_explanation },
+        { 
+          suitabilityScore: evaluation.suitability_score, 
+          taskRelevance: taskRelevance, // Use the cleaned number
+          hasExplanation: !!evaluation.detailed_explanation,
+          limitationsIdentified: evaluation.limitations?.length || 0
+        },
         evaluation.suitability_score === 'high' ? 85 : evaluation.suitability_score === 'medium' ? 70 : 50,
         Date.now() - result.processingTimeMs,
-        { toolType: tool.type }
+        { toolType: tool.type, taskRelevanceScore: taskRelevance }
       );
     }
     
@@ -963,28 +1175,31 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
 
   async processQuery(userQuery: string, mode: string): Promise<AnalysisResult> {
     const startTime = Date.now();
-    let completedTasks = 0;
+    let completeTasks = 0;
     let failedTasks = 0;
     
     this.tempAuditEntries = [];
     
-    console.log(`[AI PIPELINE] Starting ${mode} query processing with context continuity and audit trail`);
+    console.log(`[AI PIPELINE] Starting ${mode} query processing with enhanced confidence scoring`);
 
     try {
       const toolsData = await getCompressedToolsDataForAI();
-      const filteredData = await this.getIntelligentCandidates(userQuery, toolsData, mode);
       
       const context: AnalysisContext = {
         userQuery,
         mode,
-        filteredData,
+        filteredData: {}, // Will be populated by getIntelligentCandidates
         contextHistory: [],
         maxContextLength: this.maxContextTokens,
         currentContextLength: 0,
         seenToolNames: new Set<string>(),
-        auditTrail: []
+        auditTrail: [],
+        embeddingsSimilarities: new Map<string, number>()
       };
 
+      const filteredData = await this.getIntelligentCandidates(userQuery, toolsData, mode, context);
+      context.filteredData = filteredData;
+      
       this.mergeTemporaryAuditEntries(context);
 
       console.log(`[AI PIPELINE] Starting micro-tasks with ${filteredData.tools.length} tools visible`);
@@ -994,58 +1209,54 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
         { candidateTools: filteredData.tools.length, candidateConcepts: filteredData.concepts.length },
         90, 
         startTime,
-        { auditEnabled: this.auditConfig.enabled }
+        { auditEnabled: this.auditConfig.enabled, confidenceScoringEnabled: true }
       );
 
-      // MICRO-TASK SEQUENCE
+      // MICRO-TASK SEQUENCE WITH ENHANCED CONFIDENCE TRACKING
       
-      // Task 1: Scenario/Problem Analysis
       const analysisResult = await this.analyzeScenario(context);
-      if (analysisResult.success) completedTasks++; else failedTasks++;
+      if (analysisResult.success) completeTasks++; else failedTasks++;
       await this.delay(this.microTaskDelay);
 
-      // Task 2: Investigation/Solution Approach
       const approachResult = await this.generateApproach(context);
-      if (approachResult.success) completedTasks++; else failedTasks++;
+      if (approachResult.success) completeTasks++; else failedTasks++;
       await this.delay(this.microTaskDelay);
 
-      // Task 3: Critical Considerations
       const considerationsResult = await this.generateCriticalConsiderations(context);
-      if (considerationsResult.success) completedTasks++; else failedTasks++;
+      if (considerationsResult.success) completeTasks++; else failedTasks++;
       await this.delay(this.microTaskDelay);
 
-      // Task 4: Tool Selection/Evaluation (mode-dependent)
       if (mode === 'workflow') {
         const phases = toolsData.phases || [];
         for (const phase of phases) {
           const toolSelectionResult = await this.selectToolsForPhase(context, phase);
-          if (toolSelectionResult.success) completedTasks++; else failedTasks++;
+          if (toolSelectionResult.success) completeTasks++; else failedTasks++;
           await this.delay(this.microTaskDelay);
         }
       } else {
         const topTools = filteredData.tools.slice(0, 3);
         for (let i = 0; i < topTools.length; i++) {
           const evaluationResult = await this.evaluateSpecificTool(context, topTools[i], i + 1);
-          if (evaluationResult.success) completedTasks++; else failedTasks++;
+          if (evaluationResult.success) completeTasks++; else failedTasks++;
           await this.delay(this.microTaskDelay);
         }
       }
 
       const knowledgeResult = await this.selectBackgroundKnowledge(context);
-      if (knowledgeResult.success) completedTasks++; else failedTasks++;
+      if (knowledgeResult.success) completeTasks++; else failedTasks++;
       await this.delay(this.microTaskDelay);
 
       const finalResult = await this.generateFinalRecommendations(context);
-      if (finalResult.success) completedTasks++; else failedTasks++;
+      if (finalResult.success) completeTasks++; else failedTasks++;
 
       const recommendation = this.buildRecommendation(context, mode, finalResult.content);
 
       this.addAuditEntry(context, 'completion', 'pipeline-end',
-        { completedTasks, failedTasks },
+        { completedTasks: completeTasks, failedTasks },
         { finalRecommendation: !!recommendation, auditEntriesGenerated: context.auditTrail.length },
-        completedTasks > failedTasks ? 85 : 60,
+        completeTasks > failedTasks ? 85 : 60,
         startTime,
-        { totalProcessingTimeMs: Date.now() - startTime }
+        { totalProcessingTimeMs: Date.now() - startTime, confidenceScoresGenerated: context.selectedTools?.length || 0 }
       );
 
       const processingStats = {
@@ -1054,13 +1265,13 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
         finalSelectedItems: (context.selectedTools?.length || 0) + 
                            (context.backgroundKnowledge?.length || 0),
         processingTimeMs: Date.now() - startTime,
-        microTasksCompleted: completedTasks,
+        microTasksCompleted: completeTasks,
         microTasksFailed: failedTasks,
         contextContinuityUsed: true
       };
 
-      console.log(`[AI PIPELINE] Completed: ${completedTasks} tasks, Failed: ${failedTasks} tasks`);
+      console.log(`[AI PIPELINE] Completed: ${completeTasks} tasks, Failed: ${failedTasks} tasks`);
-      console.log(`[AI PIPELINE] Unique tools selected: ${context.seenToolNames.size}`);
+      console.log(`[AI PIPELINE] Enhanced confidence scores generated: ${context.selectedTools?.length || 0}`);
       console.log(`[AI PIPELINE] Audit trail entries: ${context.auditTrail.length}`);
 
       return {
@@ -1080,128 +1291,6 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
     }
   }
 
-  private calculateToolFreshness(tool: any): number {
-    // Base freshness score
-    let freshness = 70; // Default for tools without specific freshness data
-    
-    // Boost for tools with knowledge base (more maintained)
-    if (tool.knowledgebase === true) freshness += 20;
-    
-    // Boost for hosted tools (actively maintained)
-    if (isToolHosted(tool)) freshness += 15;
-    
-    // Slight boost for open source (community maintained)
-    if (tool.license && tool.license !== 'Proprietary') freshness += 5;
-    
-    return Math.min(100, freshness);
-  }
-
-  private checkDomainMatch(tool: any, userQuery: string): boolean {
-    if (!tool.domains || tool.domains.length === 0) return false;
-    
-    const queryLower = userQuery.toLowerCase();
-    
-    // Load domain keywords from environment with fallback
-    const domainKeywordsEnv = process.env.CONFIDENCE_DOMAIN_KEYWORDS || 
-      'incident-response:incident,breach,attack,compromise,response|malware-analysis:malware,virus,trojan,reverse,analysis|network-forensics:network,traffic,packet,pcap,wireshark|mobile-forensics:mobile,android,ios,phone,app|cloud-forensics:cloud,aws,azure,saas,paas';
-    
-    const domainKeywords = domainKeywordsEnv.split('|').reduce((acc, pair) => {
-      const [domain, keywords] = pair.split(':');
-      if (domain && keywords) {
-        acc[domain] = keywords.split(',');
-      }
-      return acc;
-    }, {});
-    
-    return tool.domains.some(domain => {
-      const keywords = domainKeywords[domain] || [domain.replace('-', ' ')];
-      return keywords.some(keyword => queryLower.includes(keyword));
-    });
-  }
-
-  private getMicroTaskAgreement(toolName: string, context: AnalysisContext): number {
-    // Check how many micro-tasks selected this tool
-    const microTaskEntries = context.auditTrail.filter(entry => 
-      entry.phase === 'micro-task' && 
-      entry.action.includes('selection') &&
-      entry.output && 
-      typeof entry.output === 'object' &&
-      Array.isArray(entry.output.selectedTools) &&
-      entry.output.selectedTools.includes(toolName)
-    );
-    
-    const totalMicroTasks = context.auditTrail.filter(entry => 
-      entry.phase === 'micro-task' && entry.action.includes('selection')
-    ).length;
-    
-    return totalMicroTasks > 0 ? microTaskEntries.length / totalMicroTasks : 0.8; // Default high agreement
-  }
-
-  private getEmbeddingsSimilarity(toolName: string, context: AnalysisContext): number {
-    // Extract similarity from audit trail embeddings entry
-    const embeddingsEntry = context.auditTrail.find(entry => 
-      entry.phase === 'retrieval' && entry.action === 'embeddings-search'
-    );
-    
-    if (!embeddingsEntry || !embeddingsEntry.output) return 0.5; // Default medium similarity
-    
-    // Look for similarity data in the output (implementation specific)
-    // This would need to be populated during embeddings search
-    return 0.7; // Placeholder - would need actual similarity data from embeddings
-  }
-
-  private identifyUncertaintyFactors(tool: any, context: AnalysisContext, confidence: number): string[] {
-    const factors: string[] = [];
-    
-    if (confidence < this.confidenceConfig.mediumThreshold) {
-      factors.push('Low overall confidence - consider manual validation');
-    }
-    
-    if (!this.checkDomainMatch(tool, context.userQuery)) {
-      factors.push('Domain mismatch detected - tool may not be specifically designed for this scenario');
-    }
-    
-    if (tool.skillLevel === 'expert' && /rapid|quick|urgent|triage/i.test(context.userQuery)) {
-      factors.push('Expert-level tool for rapid scenario - may be overcomplicated');
-    }
-    
-    if (tool.type === 'software' && !isToolHosted(tool) && !tool.url) {
-      factors.push('Limited access information - availability uncertain');
-    }
-    
-    if (tool.skillLevel === 'novice' && /complex|advanced|deep/i.test(context.userQuery)) {
-      factors.push('Novice-level tool for complex scenario - may lack required capabilities');
-    }
-    
-    return factors;
-  }
-
-  private identifyStrengthIndicators(tool: any, context: AnalysisContext, confidence: number): string[] {
-    const indicators: string[] = [];
-    
-    if (confidence >= this.confidenceConfig.highThreshold) {
-      indicators.push('High confidence recommendation based on multiple factors');
-    }
-    
-    if (this.checkDomainMatch(tool, context.userQuery)) {
-      indicators.push('Strong domain alignment with scenario requirements');
-    }
-    
-    if (tool.knowledgebase === true) {
-      indicators.push('Documentation and knowledge base available for guidance');
-    }
-    
-    if (isToolHosted(tool)) {
-      indicators.push('Hosted solution available for immediate access');
-    }
-    
-    if (tool.type === 'method' && /methodology|approach|process/i.test(context.userQuery)) {
-      indicators.push('Methodological approach matches procedural inquiry');
-    }
-    
-    return indicators;
-  }
-
   private buildRecommendation(context: AnalysisContext, mode: string, finalContent: string): any {
     const isWorkflow = mode === 'workflow';
     
@@ -1218,13 +1307,12 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
 
     if (isWorkflow) {
       const recommendedToolsWithConfidence = context.selectedTools?.map(st => {
-        // Calculate confidence for each tool
+        // Calculate enhanced confidence for each tool
         const confidence = this.calculateRecommendationConfidence(
           st.tool,
-          this.getEmbeddingsSimilarity(st.tool.name, context),
+          context,
-          this.checkDomainMatch(st.tool, context.userQuery),
+          st.taskRelevance || 70,
-          this.getMicroTaskAgreement(st.tool.name, context),
+          st.limitations || []
-          context
         );
         
         // Add audit entry for confidence calculation
@@ -1233,15 +1321,15 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
           { 
             overall: confidence.overall,
             components: {
-              embeddings: confidence.embeddingsQuality,
+              semantic: confidence.semanticRelevance,
-              domain: confidence.domainAlignment,
+              suitability: confidence.taskSuitability,
-              consensus: confidence.consensus,
+              consistency: confidence.methodologicalConsistency,
-              freshness: confidence.freshness
+              reliability: confidence.toolReliability
             }
           },
           confidence.overall,
           Date.now(),
-          { uncertaintyCount: confidence.uncertaintyFactors.length }
+          { uncertaintyCount: confidence.uncertaintyFactors.length, strengthCount: confidence.strengthIndicators.length }
         );
 
         return {
@@ -1264,10 +1352,9 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
       const recommendedToolsWithConfidence = context.selectedTools?.map(st => {
         const confidence = this.calculateRecommendationConfidence(
           st.tool,
-          this.getEmbeddingsSimilarity(st.tool.name, context),
+          context,
-          this.checkDomainMatch(st.tool, context.userQuery),
+          st.taskRelevance || 70,
-          this.getMicroTaskAgreement(st.tool.name, context),
+          st.limitations || []
-          context
         );
         
         this.addAuditEntry(context, 'validation', 'confidence-scoring',
@@ -1278,7 +1365,7 @@ ${JSON.stringify(conceptsToSend, null, 2)}`;
           },
           confidence.overall,
           Date.now(),
-          { strengthCount: confidence.strengthIndicators.length }
+          { strengthCount: confidence.strengthIndicators.length, limitationsCount: confidence.uncertaintyFactors.length }
         );
 
         return {