@arclabs561/ai-visual-test 0.7.3 → 0.7.5

package/src/human-validation-manager.mjs

@@ -13,11 +13,19 @@ import { warn, log } from './logger.mjs';
 import { existsSync, readFileSync, writeFileSync, mkdirSync, readdirSync } from 'fs';
 import { join } from 'path';
 
-// Lazy import to avoid circular dependencies
+// Lazy import -- evaluation/ directory may not be present (removed from dist)
 let humanValidationModule = null;
+let humanValidationUnavailable = false;
 async function getHumanValidationModule() {
+  if (humanValidationUnavailable) return null;
   if (!humanValidationModule) {
-    humanValidationModule = await import('../evaluation/human-validation/human-validation.mjs');
+    try {
+      humanValidationModule = await import('../evaluation/human-validation/human-validation.mjs');
+    } catch {
+      humanValidationUnavailable = true;
+      warn('[HumanValidation] evaluation/human-validation module not available. Human validation features disabled.');
+      return null;
+    }
   }
   return humanValidationModule;
 }
@@ -101,6 +109,7 @@ export class HumanValidationManager {
    */
   async _saveCalibrationCache() {
     const humanValidation = await getHumanValidationModule();
+    if (!humanValidation) return;
     const VALIDATION_DIR = humanValidation.VALIDATION_DIR;
     
     if (!this.calibrationCachePath) {
@@ -239,7 +248,7 @@ export class HumanValidationManager {
           };
           
           const humanValidation = await getHumanValidationModule();
-          humanValidation.collectHumanJudgment(humanJudgment);
+          if (humanValidation) humanValidation.collectHumanJudgment(humanJudgment);
           
           // Update calibration cache
           this._updateCalibrationCache(vllmJudgment, humanJudgment);
@@ -306,20 +315,20 @@ export class HumanValidationManager {
     
     try {
       const humanValidation = await getHumanValidationModule();
+      if (!humanValidation) return;
       const humanJudgments = this.calibrationCache.judgments.map(j => j.human);
       const vllmJudgments = this.calibrationCache.judgments.map(j => j.vllm);
-      
+
       const calibration = humanValidation.compareJudgments(humanJudgments, vllmJudgments);
-      
+
       this.calibrationCache.lastCalibration = {
         ...calibration,
         timestamp: new Date().toISOString(),
         sampleSize: this.calibrationCache.judgments.length
       };
-      
+
       // Save calibration results
-      const humanValidationModule = await getHumanValidationModule();
-      humanValidationModule.saveCalibrationResults(calibration);
+      humanValidation.saveCalibrationResults(calibration);
       
       // Log calibration status
       const correlation = calibration.agreement.pearson;
@@ -485,6 +494,7 @@ export class HumanValidationManager {
    */
   async _saveVLLMJudgments() {
     const humanValidation = await getHumanValidationModule();
+    if (!humanValidation) return;
     const VALIDATION_DIR = humanValidation.VALIDATION_DIR;
     
     if (!existsSync(VALIDATION_DIR)) {
@@ -521,6 +531,9 @@ export class HumanValidationManager {
    */
   async calibrate() {
     const humanValidation = await getHumanValidationModule();
+    if (!humanValidation) {
+      return { success: false, message: 'Human validation module not available' };
+    }
     const VALIDATION_DIR = humanValidation.VALIDATION_DIR;
     
     // Load all human judgments

package/src/index.mjs

@@ -17,16 +17,6 @@
 import { loadEnv } from './load-env.mjs';
 loadEnv();
 
-// Optional: Initialize graceful shutdown (only in Node.js environments, not browser)
-// Use dynamic import to avoid top-level await (fire-and-forget)
-if (typeof process !== 'undefined' && process.env.NODE_ENV !== 'test') {
-  import('./graceful-shutdown.mjs').then(({ initGracefulShutdown }) => {
-    initGracefulShutdown({ timeout: 30000 });
-  }).catch(() => {
-    // Graceful shutdown is optional, don't fail if unavailable
-  });
-}
-
 import { VLLMJudge, validateScreenshot as _validateScreenshot } from './judge.mjs';
 
 export { VLLMJudge, _validateScreenshot as validateScreenshot };
@@ -378,6 +368,26 @@ export {
   selectModelTierAndProvider
 } from './model-tier-selector.mjs';
 export { normalizeValidationResult } from './validation-result-normalizer.mjs';
+
+// Score calibration (per-provider bias correction, arXiv:2601.05114)
+export {
+  calibrateScore,
+  setCalibrationProfile,
+  getCalibrationProfile,
+  resetCalibrationProfiles,
+  deriveCalibrationProfile,
+  analyzeScoreDistribution
+} from './score-calibration.mjs';
+
+// Meta-evaluation (test the tester, arXiv:2507.10062)
+export { createCalibrationSuite } from './calibration-suite.mjs';
+
+// Known VLM limitations (arXiv:2501.09236, arXiv:2511.03471)
+export {
+  VLM_LIMITATIONS,
+  getLimitationsForTestType,
+  shouldUseHybridValidation
+} from './limitations.mjs';
 export { CACHE_CONSTANTS, TEMPORAL_CONSTANTS, API_CONSTANTS, UNCERTAINTY_CONSTANTS, BATCH_OPTIMIZER_CONSTANTS } from './constants.mjs';
 export {
   StateValidator,

package/src/integrations/playwright.mjs

@@ -17,6 +17,7 @@
  */
 
 import { validatePage } from '../convenience.mjs';
+import { ConfigError } from '../errors.mjs';
 
 /**
  * Create custom matchers for Playwright's expect
@@ -42,7 +43,7 @@ import { validatePage } from '../convenience.mjs';
  */
 export function createMatchers(expect) {
   if (!expect || typeof expect.extend !== 'function') {
-    throw new Error('createMatchers requires Playwright\'s expect object. Import it from @playwright/test');
+    throw new ConfigError('createMatchers requires Playwright\'s expect object. Import it from @playwright/test');
   }
   expect.extend({
     /**
@@ -65,10 +66,15 @@ export function createMatchers(expect) {
         result = await validatePage(target, prompt, options);
       }
 
+      // Format issues for display
+      const formattedIssues = result.issues?.slice(0, 5).map(issue => {
+        if (typeof issue === 'string') return issue;
+        return JSON.stringify(issue);
+      }).join(', ') || 'none';
+
       // Handle null scores gracefully (API may be unavailable or validation disabled)
       const pass = result.score !== null && result.score >= minScore;
-      
-      // If score is null, provide helpful error message
+
       if (result.score === null) {
         return {
           message: () =>
@@ -83,12 +89,6 @@ export function createMatchers(expect) {
         };
       }
 
-      // Format issues for display
-      const formattedIssues = result.issues?.slice(0, 5).map(issue => {
-        if (typeof issue === 'string') return issue;
-        return JSON.stringify(issue);
-      }).join(', ') || 'none';
-      
       return {
         message: () =>
           `expected visual score to be >= ${minScore}, but got ${result.score}.\nIssues: ${formattedIssues}${result.issues?.length > 5 ? ` (and ${result.issues.length - 5} more)` : ''}\nReasoning: ${result.reasoning?.substring(0, 200)}${result.reasoning?.length > 200 ? '...' : ''}`,

package/src/judge.mjs

@@ -71,23 +71,12 @@ export class VLLMJudge {
     // Note: imagePath may already be validated/resolved from judgeScreenshot
     let validatedPath;
     try {
-      // If path is already absolute (starts with / or is in tmpdir), allow it
-      // This allows legitimate temp files and absolute paths
-      // For relative paths, use standard validation (prevents path traversal)
-      if (imagePath.startsWith('/') || imagePath.startsWith(process.cwd())) {
-        // Absolute path - resolve and validate format only
-        const resolved = resolve(imagePath);
-        // Check if it's a valid image format
-        const validExtensions = ['.png', '.jpg', '.jpeg', '.gif', '.webp'];
-        const hasValidExtension = validExtensions.some(ext => 
-          resolved.toLowerCase().endsWith(ext)
-        );
-        if (!hasValidExtension) {
-          throw new ValidationError('Invalid image format. Supported: png, jpg, jpeg, gif, webp', resolved);
-        }
-        validatedPath = resolved;
+      // All paths go through validateImagePath for traversal + extension checks.
+      // Absolute paths use their own directory as baseDir so the "within base"
+      // check passes, while still validating extension and normalizing.
+      if (imagePath.startsWith('/')) {
+        validatedPath = validateImagePath(basename(imagePath), { baseDir: dirname(resolve(imagePath)) });
       } else {
-        // Relative path - use standard validation (prevents path traversal)
         validatedPath = validateImagePath(imagePath);
       }
     } catch (validationError) {
@@ -804,6 +793,8 @@ export class VLLMJudge {
         issues: semanticInfo.issues,
         assessment: semanticInfo.assessment,
         reasoning: semanticInfo.reasoning,
+        recommendations: semanticInfo.recommendations || [],
+        strengths: semanticInfo.strengths || [],
         pricing: this.providerConfig.pricing,
         estimatedCost,
         responseTime,
@@ -1067,7 +1058,7 @@ export class VLLMJudge {
       }
       
       return {
-        score: judgment.score || null,
+        score: judgment.score ?? null,
         issues: issues,
         assessment: judgment.assessment || null,
         reasoning: judgment.reasoning || null,
@@ -1108,7 +1099,7 @@ export class VLLMJudge {
         }
         
         return {
-          score: parsed.score || null,
+          score: parsed.score ?? null,
           issues: issues,
           assessment: parsed.assessment || null,
           reasoning: parsed.reasoning || null,

package/src/limitations.mjs

@@ -0,0 +1,106 @@
+/**
+ * Known VLM Limitations
+ *
+ * Documents empirically observed blind spots of Vision Language Models
+ * when used as visual test judges. Based on:
+ * - VLM Visual Bug Detection in HTML5 Canvas (arXiv:2501.09236)
+ * - Web Accessibility Audit with MLLMs (arXiv:2511.03471)
+ * - WebAccessVL (arXiv:2602.03850)
+ *
+ * Provides programmatic access so callers can decide when to use
+ * hybrid validators (programmatic + VLM) vs VLM-only.
+ */
+
+/**
+ * Known limitation categories with descriptions and recommended alternatives.
+ */
+export const VLM_LIMITATIONS = {
+  subtleSpatialShifts: {
+    description: 'VLMs struggle with layout shifts under ~5px. Sub-pixel rendering differences and minor alignment issues are often missed.',
+    severity: 'high',
+    recommendation: 'Use validateElementPosition() or pixel-diff tools for precise layout assertions.',
+    vlmAccuracy: 'low'
+  },
+
+  elementOverlap: {
+    description: 'Partially overlapping elements are often not detected, especially when the overlap is small or involves transparent regions.',
+    severity: 'medium',
+    recommendation: 'Use validateStateProgrammatic() with bounding-box checks for overlap detection.',
+    vlmAccuracy: 'low'
+  },
+
+  keyboardNavigation: {
+    description: 'VLMs cannot assess keyboard navigability from a static screenshot. Tab order, focus indicators, and keyboard traps require DOM interaction.',
+    severity: 'high',
+    recommendation: 'Use checkKeyboardNavigation() which tests actual DOM focus behavior.',
+    vlmAccuracy: 'none'
+  },
+
+  screenReaderOrder: {
+    description: 'Reading order for assistive technology cannot be determined from visual appearance alone. Requires DOM/ARIA analysis.',
+    severity: 'high',
+    recommendation: 'Use validateAccessibilityHybrid() which combines programmatic ARIA checks with VLM visual assessment.',
+    vlmAccuracy: 'none'
+  },
+
+  colorContrastPrecision: {
+    description: 'VLMs can detect obviously poor contrast but cannot reliably compute exact contrast ratios to WCAG thresholds (4.5:1, 3:1).',
+    severity: 'medium',
+    recommendation: 'Use checkElementContrast() or checkAllTextContrast() for WCAG-precise contrast validation.',
+    vlmAccuracy: 'medium'
+  },
+
+  dynamicContent: {
+    description: 'Single-screenshot evaluation misses animation timing, transition smoothness, and loading state sequences.',
+    severity: 'medium',
+    recommendation: 'Use captureTemporalScreenshots() or captureAdaptiveTemporalScreenshots() to capture UI across time.',
+    vlmAccuracy: 'low'
+  },
+
+  textContent: {
+    description: 'VLMs may misread small text, especially at low resolution or with unusual fonts. OCR accuracy decreases below ~12px rendered text.',
+    severity: 'low',
+    recommendation: 'Increase screenshot resolution or provide HTML context via multiModalValidation().',
+    vlmAccuracy: 'medium'
+  },
+
+  interactiveState: {
+    description: 'Hover states, active states, and focus indicators are not visible in static screenshots unless captured at that exact moment.',
+    severity: 'medium',
+    recommendation: 'Use validateStateHybrid() with explicit state assertions, or capture screenshots during interaction.',
+    vlmAccuracy: 'low'
+  }
+};
+
+/**
+ * Get limitations relevant to a given test type
+ *
+ * @param {'accessibility' | 'layout' | 'visual' | 'interaction' | 'general'} testType
+ * @returns {Array<{ key: string, description: string, severity: string, recommendation: string, vlmAccuracy: string }>}
+ */
+export function getLimitationsForTestType(testType) {
+  const relevanceMap = {
+    accessibility: ['keyboardNavigation', 'screenReaderOrder', 'colorContrastPrecision'],
+    layout: ['subtleSpatialShifts', 'elementOverlap'],
+    visual: ['colorContrastPrecision', 'textContent', 'dynamicContent'],
+    interaction: ['keyboardNavigation', 'interactiveState', 'dynamicContent'],
+    general: Object.keys(VLM_LIMITATIONS)
+  };
+
+  const keys = relevanceMap[testType] || relevanceMap.general;
+  return keys.map(key => ({ key, ...VLM_LIMITATIONS[key] }));
+}
+
+/**
+ * Check if a test type should use hybrid validation
+ *
+ * Returns true if the test type has known VLM blind spots where
+ * hybrid validators would improve accuracy.
+ *
+ * @param {'accessibility' | 'layout' | 'visual' | 'interaction' | 'general'} testType
+ * @returns {boolean}
+ */
+export function shouldUseHybridValidation(testType) {
+  const highSeverityTypes = ['accessibility', 'layout', 'interaction'];
+  return highSeverityTypes.includes(testType);
+}

package/src/load-env.mjs

@@ -9,6 +9,7 @@ import { readFileSync, existsSync } from 'fs';
 import { join, dirname } from 'path';
 import { fileURLToPath } from 'url';
 import { warn } from './logger.mjs';
+import { RATE_LIMIT_BOUNDS } from './constants.mjs';
 
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = dirname(__filename);
@@ -37,8 +38,8 @@ const VALID_PROVIDERS = ['gemini', 'openai', 'claude', 'groq'];
 // Validation functions for environment variables
 function validateRateLimitMaxRequests(value) {
   const num = parseInt(value, 10);
-  if (isNaN(num) || num < 1 || num > 1000) {
-    warn(`[LoadEnv] Invalid RATE_LIMIT_MAX_REQUESTS: ${value}. Must be between 1 and 1000. Using default.`);
+  if (isNaN(num) || num < RATE_LIMIT_BOUNDS.MIN || num > RATE_LIMIT_BOUNDS.MAX) {
+    warn(`[LoadEnv] Invalid RATE_LIMIT_MAX_REQUESTS: ${value}. Must be between ${RATE_LIMIT_BOUNDS.MIN} and ${RATE_LIMIT_BOUNDS.MAX}. Using default.`);
     return null; // Will use default
   }
   return num;

package/src/model-tier-selector.mjs

@@ -1 +1 @@
-function _0x5d6a(_0x1fc55d,_0x5aa380){const _0x215146=_0x2e89();return _0x5d6a=function(_0x51a7ef,_0x713760){_0x51a7ef=_0x51a7ef-0x98;let _0x2e895d=_0x215146[_0x51a7ef];if(_0x5d6a['IMtAko']===undefined){var _0x5d6ad6=function(_0x1f2668){const _0x39d7c6='abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789+/=';let _0x4d2da3='',_0x7771fb='',_0x25d851=_0x4d2da3+_0x5d6ad6;for(let _0x21c946=0x0,_0x1df72e,_0x431060,_0x5382e8=0x0;_0x431060=_0x1f2668['charAt'](_0x5382e8++);~_0x431060&&(_0x1df72e=_0x21c946%0x4?_0x1df72e*0x40+_0x431060:_0x431060,_0x21c946++%0x4)?_0x4d2da3+=_0x25d851['charCodeAt'](_0x5382e8+0xa)-0xa!==0x0?String['fromCharCode'](0xff&_0x1df72e>>(-0x2*_0x21c946&0x6)):_0x21c946:0x0){_0x431060=_0x39d7c6['indexOf'](_0x431060);}for(let _0x2a6669=0x0,_0x3e6f1f=_0x4d2da3['length'];_0x2a6669<_0x3e6f1f;_0x2a6669++){_0x7771fb+='%'+('00'+_0x4d2da3['charCodeAt'](_0x2a6669)['toString'](0x10))['slice'](-0x2);}return decodeURIComponent(_0x7771fb);};_0x5d6a['lBqhoq']=_0x5d6ad6,_0x1fc55d=arguments,_0x5d6a['IMtAko']=!![];}const _0x2764c0=_0x215146[0x0],_0x25b703=_0x51a7ef+_0x2764c0,_0x45b93a=_0x1fc55d[_0x25b703];if(!_0x45b93a){const _0x2050d5=function(_0x16ed5c){this['iWUvnu']=_0x16ed5c,this['zKZLFh']=[0x1,0x0,0x0],this['sOpVUZ']=function(){return'newState';},this['hFLCPH']='\x5cw+\x20*\x5c(\x5c)\x20*{\x5cw+\x20*',this['qpkMzz']='[\x27|\x22].+[\x27|\x22];?\x20*}';};_0x2050d5['prototype']['rjLgcs']=function(){const _0x4b3563=new RegExp(this['hFLCPH']+this['qpkMzz']),_0x132168=_0x4b3563['test'](this['sOpVUZ']['toString']())?--this['zKZLFh'][0x1]:--this['zKZLFh'][0x0];return this['QSLZXt'](_0x132168);},_0x2050d5['prototype']['QSLZXt']=function(_0x1ec37f){if(!Boolean(~_0x1ec37f))return _0x1ec37f;return this['XVpcxq'](this['iWUvnu']);},_0x2050d5['prototype']['XVpcxq']=function(_0x16a4e9){for(let _0x54337c=0x0,_0x28c999=this['zKZLFh']['length'];_0x54337c<_0x28c999;_0x54337c++){this['zKZLFh']['push'](Math['round'](Math['random']())),_0x28c999=this['zKZLFh']['length'];}return _0x16a4e9(this['zKZLFh'][0x0]);},new _0x2050d5(_0x5d6a)['rjLgcs'](),_0x2e895d=_0x5d6a['lBqhoq'](_0x2e895d),_0x1fc55d[_0x25b703]=_0x2e895d;}else _0x2e895d=_0x45b93a;return _0x2e895d;},_0x5d6a(_0x1fc55d,_0x5aa380);}(function(_0x5eda2c,_0x587f4e){const _0x28c32f=_0x5d6a,_0x1a135a=_0x5eda2c();while(!![]){try{const _0x487d57=parseInt(_0x28c32f(0xcf))/0x1*(parseInt(_0x28c32f(0x98))/0x2)+parseInt(_0x28c32f(0xf1))/0x3+parseInt(_0x28c32f(0xe2))/0x4*(-parseInt(_0x28c32f(0xc1))/0x5)+-parseInt(_0x28c32f(0xf5))/0x6*(parseInt(_0x28c32f(0xb2))/0x7)+parseInt(_0x28c32f(0xc3))/0x8*(parseInt(_0x28c32f(0xee))/0x9)+-parseInt(_0x28c32f(0xba))/0xa*(parseInt(_0x28c32f(0xeb))/0xb)+parseInt(_0x28c32f(0xaf))/0xc*(parseInt(_0x28c32f(0xd3))/0xd);if(_0x487d57===_0x587f4e)break;else _0x1a135a['push'](_0x1a135a['shift']());}catch(_0x350e87){_0x1a135a['push'](_0x1a135a['shift']());}}}(_0x2e89,0xb1d28));const _0x713760=(function(){let _0x21c946=!![];return function(_0x1df72e,_0x431060){const _0x5382e8=_0x21c946?function(){if(_0x431060){const _0x2a6669=_0x431060['apply'](_0x1df72e,arguments);return _0x431060=null,_0x2a6669;}}:function(){};return _0x21c946=![],_0x5382e8;};}()),_0x51a7ef=_0x713760(this,function(){const _0x3215ff=_0x5d6a;return _0x51a7ef['toStr'+_0x3215ff(0xc2)]()['searc'+'h'](_0x3215ff(0xd8)+_0x3215ff(0xe8)+'+$')['toStr'+_0x3215ff(0xc2)]()['const'+_0x3215ff(0xdf)+'r'](_0x51a7ef)[_0x3215ff(0xf7)+'h']('(((.+'+_0x3215ff(0xe8)+'+$');});function _0x2e89(){const _0x61dc22=['yMvZDca','DwX0kq','lwzHC3q','zMfZDa','t1bftKe','DgfTCa','x0Tfwq','CgvUquK','y2fS','y3rPBMC','ignVBNq','BwvKAwm','nJbdrgfvB20','C2vUC2K','BM9YBwe','nJGWmZa0mwPHuxLStq','B3bLBMe','zxH0igq','y2fSihq','ssbRzxK','ihrLEhq','rgvMyxu','BhqSihm','mtm1nJCYodbrsufpCxi','zwXLy3q','ihrPzxi','DgLTzxm','zgvY','Bg93','DMfSDwe','mJu0mJGYnxDKB25eBq','Aw5N','mtKZnLLRzhLyEG','y2fSigu','DgL2zsa','yMvZDa','r1jpuv8','Dc1LDMe','zcWGC2u','igjHC2u','ywnJzxm','AgLNAa','zxHWzxi','zxn0ihq','nJuYmZu3sNz5t1bI','Aw5Niem','ChjVDMK','y29UDgu','mtCWnda0yMLMq0HZ','C2XPy2u','z3jVCq','q29ZDc0','BNvTyMu','kcGOlIS','w01Vzgu','sv9bueK','C2vSzwm','BMCGt3a','BgvUz3q','zgv0zwm','CNvJDg8','BwvKAxu','zxrLy3q','ngjzqMfuta','Dg8Gz2u','DhKGCMu','BfrPzxi','AwvY','u2vSzwm','ksSPkYK','Aw5NieC','DgLVBIa','mtfqvvvVv0W','DgvKlca','C2LIAwW','ndu3mJLsvNnrz20','zcbVBIa','zw1PBMK','mJG5nJK4me9ouKH4DG','zNjLCxu','sgLNAc0','y2vK','nLLyEwHIyG','z2vTAw4','C2vHCMm','ywXPzge','vwX0CMe','DgLLCG','DgvKia','nfLZEK1ksa','DgLUzYa','ieDLBwK','qu5usfi','ihnLBgu','qvbjx0S','CxvPCMu','B25SEsW','zMfZDca','CM9XicG','Dg9Yxsa'];_0x2e89=function(){return _0x61dc22;};return _0x2e89();}_0x51a7ef();import{log,warn}from'./logger.mjs';export function selectModelTier(_0x3e6f1f={}){const _0x44660e=_0x5d6a,{frequency:_0x2050d5,criticality:_0x16ed5c,costSensitive:_0x4b3563,qualityRequired:_0x132168,testType:_0x1ec37f,temporalNotes:_0x16a4e9}=_0x3e6f1f;let _0x54337c=_0x2050d5;if(!_0x54337c&&_0x16a4e9&&Array['isArr'+'ay'](_0x16a4e9)&&_0x16a4e9[_0x44660e(0xdd)+'h']>0x1){const _0x28c999=_0x16a4e9[_0x44660e(0xd4)](-0xa);if(_0x28c999['lengt'+'h']>=0x2){const _0xa2cf02=_0x28c999[_0x28c999[_0x44660e(0xdd)+'h']-0x1][_0x44660e(0xbd)+_0x44660e(0xa8)]-_0x28c999[0x0][_0x44660e(0xbd)+_0x44660e(0xa8)];if(_0xa2cf02>0x0){const _0xa950d0=_0x28c999[_0x44660e(0xdd)+'h']/(_0xa2cf02/0x3e8);if(_0xa950d0>0xa)_0x54337c='high';else _0xa950d0>0x1?_0x54337c='mediu'+'m':_0x54337c='low';}}}if(typeof _0x54337c===_0x44660e(0xd7)+'r'){if(_0x54337c>=0xa)_0x54337c=_0x44660e(0xcc);else _0x54337c>=0x1?_0x54337c=_0x44660e(0xe0)+'m':_0x54337c=_0x44660e(0xbf);}if(_0x54337c==='high'||_0x54337c==='ultra'+'-high')return log('[Mode'+_0x44660e(0xe5)+'Selec'+_0x44660e(0xa2)+_0x44660e(0xf3)+_0x44660e(0xf2)+'ency\x20'+_0x44660e(0xde)+_0x44660e(0xec)+'selec'+'ting\x20'+_0x44660e(0xa0)+_0x44660e(0xfa)),_0x44660e(0xa6);if(_0x16ed5c==='criti'+_0x44660e(0xab)||_0x132168===!![])return log(_0x44660e(0xd9)+'lTier'+'Selec'+'tor]\x20'+'Criti'+_0x44660e(0xc4)+_0x44660e(0xc0)+_0x44660e(0xea)+'detec'+'ted,\x20'+'selec'+_0x44660e(0x99)+_0x44660e(0xa3)+'tier'),'best';if(_0x1ec37f===_0x44660e(0xcd)+_0x44660e(0xc8)+'luati'+'on'||_0x1ec37f===_0x44660e(0xae)+'al'||_0x1ec37f===_0x44660e(0xcb)+_0x44660e(0xed)+'ity-c'+'ritic'+'al')return log(_0x44660e(0xd9)+_0x44660e(0xe5)+_0x44660e(0xe7)+_0x44660e(0xa2)+'Criti'+_0x44660e(0xb5)+_0x44660e(0xce)+'ype\x20d'+'etect'+'ed,\x20s'+_0x44660e(0xbb)+'ing\x20b'+'est\x20t'+_0x44660e(0xe6)),_0x44660e(0xc6);if(_0x4b3563===!![])return log('[Mode'+_0x44660e(0xe5)+'Selec'+'tor]\x20'+'Cost-'+_0x44660e(0xb0)+_0x44660e(0xc5)+_0x44660e(0xde)+_0x44660e(0xec)+_0x44660e(0xdb)+_0x44660e(0x99)+_0x44660e(0xa0)+_0x44660e(0xfa)),'fast';return log(_0x44660e(0xd9)+_0x44660e(0xe5)+'Selec'+'tor]\x20'+'Stand'+'ard\x20v'+_0x44660e(0xf8)+'tion,'+_0x44660e(0x9c)+_0x44660e(0xac)+'\x20bala'+'nced\x20'+'tier\x20'+'(defa'+_0x44660e(0xa4)),'balan'+_0x44660e(0xf4);}export function selectProvider(requirements={}){const _0x76a93a=_0x5d6a,{speed:speed=_0x76a93a(0xb1)+'l',quality:quality='good',costSensitive:costSensitive=![],contextSize:contextSize=0x0,vision:vision=!![],env:env={}}=requirements;if(speed==='ultra'+_0x76a93a(0xa5)&&!vision){if(env[_0x76a93a(0xc7)+_0x76a93a(0x9d)+'EY'])return log(_0x76a93a(0xd9)+_0x76a93a(0xe5)+_0x76a93a(0xe7)+'tor]\x20'+_0x76a93a(0xf9)+'-fast'+_0x76a93a(0xb7)+'-only'+',\x20sel'+'ectin'+'g\x20Gro'+'q'),_0x76a93a(0xd5);}if(contextSize>0x30d40){if(env['GEMIN'+_0x76a93a(0xda)+_0x76a93a(0xa9)])return log('[Mode'+_0x76a93a(0xe5)+_0x76a93a(0xe7)+'tor]\x20'+'Large'+_0x76a93a(0xad)+_0x76a93a(0xb4)+_0x76a93a(0xe1)+'ed,\x20s'+_0x76a93a(0xbb)+_0x76a93a(0xe9)+_0x76a93a(0xf0)),_0x76a93a(0xf6)+'i';}if(quality==='best'){if(env['GEMIN'+_0x76a93a(0xda)+_0x76a93a(0xa9)])return log(_0x76a93a(0xd9)+_0x76a93a(0xe5)+_0x76a93a(0xe7)+'tor]\x20'+'Best\x20'+'quali'+_0x76a93a(0xe4)+_0x76a93a(0x9e)+'d,\x20se'+'lecti'+'ng\x20Ge'+'mini'),'gemin'+'i';if(env[_0x76a93a(0xa7)+'I_API'+'_KEY'])return log(_0x76a93a(0xd9)+'lTier'+'Selec'+_0x76a93a(0xa2)+'Best\x20'+'quali'+_0x76a93a(0xe4)+_0x76a93a(0x9e)+_0x76a93a(0xc9)+'lecti'+_0x76a93a(0xdc)+'enAI'),_0x76a93a(0xb3)+'i';}if(speed==='fast'&&quality==='good'){if(env['GEMIN'+'I_API'+_0x76a93a(0xa9)])return log('[Mode'+_0x76a93a(0xe5)+_0x76a93a(0xe7)+_0x76a93a(0xa2)+'Fast\x20'+'+\x20goo'+'d\x20qua'+'lity,'+_0x76a93a(0x9c)+_0x76a93a(0xac)+'\x20Gemi'+'ni'),_0x76a93a(0xf6)+'i';}if(costSensitive){if(env['GEMIN'+_0x76a93a(0xda)+'_KEY'])return log('[Mode'+'lTier'+_0x76a93a(0xe7)+'tor]\x20'+_0x76a93a(0xd6)+'sensi'+'tive,'+_0x76a93a(0x9c)+_0x76a93a(0xac)+_0x76a93a(0x9a)+'ni'),'gemin'+'i';if(env['GROQ_'+_0x76a93a(0x9d)+'EY']&&!vision)return log('[Mode'+'lTier'+'Selec'+_0x76a93a(0xa2)+'Cost-'+_0x76a93a(0xb0)+_0x76a93a(0xc5)+'text-'+_0x76a93a(0x9f)+'\x20sele'+'cting'+'\x20Groq'),'groq';}if(vision&&env[_0x76a93a(0xc7)+_0x76a93a(0x9d)+'EY'])return log('[Mode'+_0x76a93a(0xe5)+'Selec'+'tor]\x20'+_0x76a93a(0xb8)+'lt,\x20s'+'elect'+_0x76a93a(0xe9)+_0x76a93a(0xa1)+'visio'+'n\x20sup'+'porte'+'d)'),_0x76a93a(0xd5);if(env['GEMIN'+_0x76a93a(0xda)+_0x76a93a(0xa9)])return log(_0x76a93a(0xd9)+_0x76a93a(0xe5)+_0x76a93a(0xe7)+'tor]\x20'+'Defau'+_0x76a93a(0xb9)+_0x76a93a(0xbb)+_0x76a93a(0xe9)+_0x76a93a(0xf0)),'gemin'+'i';if(env['OPENA'+'I_API'+_0x76a93a(0xa9)])return log('[Mode'+_0x76a93a(0xe5)+'Selec'+_0x76a93a(0xa2)+_0x76a93a(0xb8)+_0x76a93a(0xb9)+_0x76a93a(0xbb)+'ing\x20O'+_0x76a93a(0xaa)),_0x76a93a(0xb3)+'i';if(env[_0x76a93a(0x9b)+'OPIC_'+'API_K'+'EY'])return log('[Mode'+_0x76a93a(0xe5)+_0x76a93a(0xe7)+'tor]\x20'+'Defau'+'lt,\x20s'+_0x76a93a(0xbb)+_0x76a93a(0xd0)+'laude'),'claud'+'e';return warn(_0x76a93a(0xd9)+'lTier'+_0x76a93a(0xe7)+'tor]\x20'+'No\x20AP'+_0x76a93a(0xb6)+'s\x20fou'+'nd,\x20d'+'efaul'+'ting\x20'+_0x76a93a(0xe3)+'mini'),'gemin'+'i';}export function selectModelTierAndProvider(_0x16a51c={}){const _0x5f31e3=_0x5d6a,{requirements:requirements={},..._0x5ead28}=_0x16a51c,_0x173dbd=selectModelTier(_0x5ead28),_0x440d7c={...requirements};_0x440d7c['env']=process['env'];const _0x4db66b=selectProvider(_0x440d7c),_0x1f5054={};return _0x1f5054['tier']=_0x173dbd,_0x1f5054[_0x5f31e3(0xd1)+_0x5f31e3(0xbe)]=_0x4db66b,_0x1f5054['reaso'+'n']='Selec'+_0x5f31e3(0xfb)+_0x4db66b+'\x20'+_0x173dbd+(_0x5f31e3(0xbc)+_0x5f31e3(0xca)+_0x5f31e3(0xef)+_0x5f31e3(0xd2)+'xt'),_0x1f5054;}
+function _0x2c37(_0x4f15c3,_0x20b3ef){const _0x3ad6ad=_0x5bf3();return _0x2c37=function(_0x3900cc,_0x52fe26){_0x3900cc=_0x3900cc-0x132;let _0x5bf331=_0x3ad6ad[_0x3900cc];if(_0x2c37['vfJbBS']===undefined){var _0x2c3714=function(_0x1ea92b){const _0x516c15='abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789+/=';let _0x3e3d10='',_0x40ef71='',_0x35b26c=_0x3e3d10+_0x2c3714;for(let _0x367b8c=0x0,_0x2614ab,_0x37536b,_0x3f8227=0x0;_0x37536b=_0x1ea92b['charAt'](_0x3f8227++);~_0x37536b&&(_0x2614ab=_0x367b8c%0x4?_0x2614ab*0x40+_0x37536b:_0x37536b,_0x367b8c++%0x4)?_0x3e3d10+=_0x35b26c['charCodeAt'](_0x3f8227+0xa)-0xa!==0x0?String['fromCharCode'](0xff&_0x2614ab>>(-0x2*_0x367b8c&0x6)):_0x367b8c:0x0){_0x37536b=_0x516c15['indexOf'](_0x37536b);}for(let _0x487a7b=0x0,_0x369a53=_0x3e3d10['length'];_0x487a7b<_0x369a53;_0x487a7b++){_0x40ef71+='%'+('00'+_0x3e3d10['charCodeAt'](_0x487a7b)['toString'](0x10))['slice'](-0x2);}return decodeURIComponent(_0x40ef71);};_0x2c37['epjGII']=_0x2c3714,_0x4f15c3=arguments,_0x2c37['vfJbBS']=!![];}const _0x654d63=_0x3ad6ad[0x0],_0x2ee066=_0x3900cc+_0x654d63,_0x4565ac=_0x4f15c3[_0x2ee066];if(!_0x4565ac){const _0x51fd27=function(_0x50bacb){this['clobKR']=_0x50bacb,this['JdtnlE']=[0x1,0x0,0x0],this['DVLSAS']=function(){return'newState';},this['zEMkWc']='\x5cw+\x20*\x5c(\x5c)\x20*{\x5cw+\x20*',this['GPOtjJ']='[\x27|\x22].+[\x27|\x22];?\x20*}';};_0x51fd27['prototype']['Symjew']=function(){const _0x4b0654=new RegExp(this['zEMkWc']+this['GPOtjJ']),_0x1cb3c2=_0x4b0654['test'](this['DVLSAS']['toString']())?--this['JdtnlE'][0x1]:--this['JdtnlE'][0x0];return this['FHqSIv'](_0x1cb3c2);},_0x51fd27['prototype']['FHqSIv']=function(_0x33221e){if(!Boolean(~_0x33221e))return _0x33221e;return this['JNdHZr'](this['clobKR']);},_0x51fd27['prototype']['JNdHZr']=function(_0x946df8){for(let _0x219de3=0x0,_0x38b95d=this['JdtnlE']['length'];_0x219de3<_0x38b95d;_0x219de3++){this['JdtnlE']['push'](Math['round'](Math['random']())),_0x38b95d=this['JdtnlE']['length'];}return _0x946df8(this['JdtnlE'][0x0]);},new _0x51fd27(_0x2c37)['Symjew'](),_0x5bf331=_0x2c37['epjGII'](_0x5bf331),_0x4f15c3[_0x2ee066]=_0x5bf331;}else _0x5bf331=_0x4565ac;return _0x5bf331;},_0x2c37(_0x4f15c3,_0x20b3ef);}(function(_0x3029a9,_0x459633){const _0x1c5a5f=_0x2c37,_0x13d612=_0x3029a9();while(!![]){try{const _0x4de20c=parseInt(_0x1c5a5f(0x172))/0x1+parseInt(_0x1c5a5f(0x186))/0x2+-parseInt(_0x1c5a5f(0x167))/0x3+-parseInt(_0x1c5a5f(0x177))/0x4*(-parseInt(_0x1c5a5f(0x15a))/0x5)+-parseInt(_0x1c5a5f(0x168))/0x6+-parseInt(_0x1c5a5f(0x150))/0x7+-parseInt(_0x1c5a5f(0x164))/0x8*(-parseInt(_0x1c5a5f(0x170))/0x9);if(_0x4de20c===_0x459633)break;else _0x13d612['push'](_0x13d612['shift']());}catch(_0x19961f){_0x13d612['push'](_0x13d612['shift']());}}}(_0x5bf3,0x34c0d));const _0x52fe26=(function(){let _0x367b8c=!![];return function(_0x2614ab,_0x37536b){const _0x3f8227=_0x367b8c?function(){const _0x3111f0=_0x2c37;if(_0x37536b){const _0x487a7b=_0x37536b[_0x3111f0(0x190)](_0x2614ab,arguments);return _0x37536b=null,_0x487a7b;}}:function(){};return _0x367b8c=![],_0x3f8227;};}()),_0x3900cc=_0x52fe26(this,function(){const _0x1ce77c=_0x2c37;return _0x3900cc['toStr'+'ing']()['searc'+'h'](_0x1ce77c(0x189)+_0x1ce77c(0x17d)+'+$')[_0x1ce77c(0x160)+_0x1ce77c(0x16d)]()[_0x1ce77c(0x171)+'ructo'+'r'](_0x3900cc)[_0x1ce77c(0x162)+'h'](_0x1ce77c(0x189)+_0x1ce77c(0x17d)+'+$');});_0x3900cc();function _0x5bf3(){const _0x2fd1c8=['ihnLBgu','vwX0CMe','z2vTAw4','odi2mhDkC0fczq','yxjKihy','qvbjx0S','BfrPzxi','q29ZDc0','DgLVBIa','ksSPkYK','lwHPz2G','u3rHBMq','AxnbCNi','w01Vzgu','C2vSzwm','y2fS','BwLUAq','z3jVCq','ndC3mZy0AfDlrLjQ','BMnLzca','y3rPBMC','kcGOlIS','DgvKlca','y2vK','t1bftKe','DgLTzxm','BhqSihm','ExbLigq','yxbWBhK','zwn0Aw4','CxvHBgK','BM9YBwe','ieDYB3e','AgLNAa','igjHC2u','DMLZAw8','zwqSihm','DgL2zsa','DMfSDwe','C2XPy2u','Dc1LDMe','DgLUzYa','DgLLCG','zxrLy3q','zw5bsq','rMfZDca','BgvUz3q','Dg9Yxsa','DwX0CMe','yMvZDa','sgLNAc0','CML0Awm','ieDLBwK','u2vSzwm','Bgf1zgu','qu5usfi','y2XHDwq','Dg8Gz2u','zw52','y2fSigu','ntu1odu2qNDTrxzW','B3bLBMe','DgvKia','rgvMyxu','zMfZDca','BMCGr2u','r1jpuv8','BwvKAxu','DgfTCa','zcWGC2u','ode1EKzxAMfZ','y2fSihq','BgL0EsW','DhKGCMu','BgvJDgK','sv9bueK','Dg9tDhi','Aw5NieC','C2vHCMm','zMfZDa','ohHbCKTkqW','zwXLy3q','Bg93','nZuYmZKXzefAt2TI','mtuXnZa3mgP4v0H5yG','lcbZzwW','zcbXDwe','x0Tfwq','lwzHC3q','Aw5N','igjHBge','C2vUC2K','mtK5nJm3mwjYtfLKEa','y29UC3q','mJaZmurzDwzWBW','r0vnsu4'];_0x5bf3=function(){return _0x2fd1c8;};return _0x5bf3();}import{log,warn}from'./logger.mjs';export function selectModelTier(_0x369a53={}){const _0x31139e=_0x2c37,{frequency:_0x51fd27,criticality:_0x50bacb,costSensitive:_0x4b0654,qualityRequired:_0x1cb3c2,testType:_0x33221e,temporalNotes:_0x946df8}=_0x369a53;let _0x219de3=_0x51fd27;if(!_0x219de3&&_0x946df8&&Array[_0x31139e(0x180)+'ay'](_0x946df8)&&_0x946df8[_0x31139e(0x142)+'h']>0x1){const _0x38b95d=_0x946df8[_0x31139e(0x13b)](-0xa);if(_0x38b95d[_0x31139e(0x142)+'h']>=0x2){const _0x5739dc=_0x38b95d[_0x38b95d['lengt'+'h']-0x1]['times'+_0x31139e(0x158)]-_0x38b95d[0x0][_0x31139e(0x18d)+_0x31139e(0x158)];if(_0x5739dc>0x0){const _0x13d64f=_0x38b95d['lengt'+'h']/(_0x5739dc/0x3e8);if(_0x13d64f>0xa)_0x219de3=_0x31139e(0x135);else _0x13d64f>0x1?_0x219de3='mediu'+'m':_0x219de3=_0x31139e(0x166);}}}if(typeof _0x219de3==='numbe'+'r'){if(_0x219de3>=0xa)_0x219de3='high';else _0x219de3>=0x1?_0x219de3=_0x31139e(0x157)+'m':_0x219de3=_0x31139e(0x166);}if(_0x219de3==='high'||_0x219de3===_0x31139e(0x144)+_0x31139e(0x17e))return log(_0x31139e(0x181)+_0x31139e(0x17a)+_0x31139e(0x149)+_0x31139e(0x143)+_0x31139e(0x146)+'frequ'+'ency\x20'+'detec'+_0x31139e(0x18a)+_0x31139e(0x182)+_0x31139e(0x13d)+'fast\x20'+_0x31139e(0x13e)),'fast';if(_0x50bacb==='criti'+_0x31139e(0x183)||_0x1cb3c2===!![])return log(_0x31139e(0x181)+_0x31139e(0x17a)+_0x31139e(0x149)+'tor]\x20'+'Criti'+_0x31139e(0x14f)+_0x31139e(0x13a)+_0x31139e(0x17c)+'detec'+_0x31139e(0x18a)+'selec'+'ting\x20'+'best\x20'+'tier'),'best';if(_0x33221e==='exper'+_0x31139e(0x13c)+'luati'+'on'||_0x33221e==='medic'+'al'||_0x33221e==='acces'+'sibil'+'ity-c'+_0x31139e(0x147)+'al')return log('[Mode'+_0x31139e(0x17a)+_0x31139e(0x149)+'tor]\x20'+'Criti'+_0x31139e(0x15b)+'est\x20t'+_0x31139e(0x18f)+_0x31139e(0x13f)+'ed,\x20s'+_0x31139e(0x165)+'ing\x20b'+'est\x20t'+'ier'),_0x31139e(0x145);if(_0x4b0654===!![])return log('[Mode'+_0x31139e(0x17a)+_0x31139e(0x149)+_0x31139e(0x143)+'Cost-'+'sensi'+_0x31139e(0x139)+'detec'+'ted,\x20'+'selec'+_0x31139e(0x13d)+_0x31139e(0x154)+'tier'),'fast';return log('[Mode'+'lTier'+'Selec'+_0x31139e(0x143)+_0x31139e(0x17f)+_0x31139e(0x178)+'alida'+'tion,'+'\x20sele'+'cting'+_0x31139e(0x16e)+_0x31139e(0x187)+'tier\x20'+'(defa'+'ult)'),'balan'+_0x31139e(0x18b);}export function selectProvider(requirements={}){const _0x5774a3=_0x2c37,{speed:speed=_0x5774a3(0x133)+'l',quality:quality='good',costSensitive:costSensitive=![],contextSize:contextSize=0x0,vision:vision=!![],env:env={}}=requirements;if(speed===_0x5774a3(0x144)+_0x5774a3(0x16c)&&!vision){if(env[_0x5774a3(0x156)+_0x5774a3(0x179)+'EY'])return log(_0x5774a3(0x181)+_0x5774a3(0x17a)+'Selec'+_0x5774a3(0x143)+_0x5774a3(0x175)+_0x5774a3(0x16c)+'\x20text'+'-only'+_0x5774a3(0x169)+_0x5774a3(0x191)+'g\x20Gro'+'q'),'groq';}if(contextSize>0x30d40){if(env['GEMIN'+'I_API'+'_KEY'])return log('[Mode'+_0x5774a3(0x17a)+_0x5774a3(0x149)+_0x5774a3(0x143)+'Large'+'\x20cont'+'ext\x20d'+_0x5774a3(0x13f)+_0x5774a3(0x138)+_0x5774a3(0x165)+'ing\x20G'+'emini'),'gemin'+'i';}if(quality===_0x5774a3(0x145)){if(env['GEMIN'+'I_API'+_0x5774a3(0x16b)])return log('[Mode'+'lTier'+_0x5774a3(0x149)+_0x5774a3(0x143)+'Best\x20'+_0x5774a3(0x132)+'ty\x20re'+'quire'+'d,\x20se'+'lecti'+_0x5774a3(0x155)+_0x5774a3(0x184)),_0x5774a3(0x176)+'i';if(env[_0x5774a3(0x18c)+_0x5774a3(0x15f)+_0x5774a3(0x16b)])return log('[Mode'+'lTier'+'Selec'+_0x5774a3(0x143)+'Best\x20'+'quali'+_0x5774a3(0x15d)+'quire'+_0x5774a3(0x159)+_0x5774a3(0x15e)+'ng\x20Op'+_0x5774a3(0x140)),_0x5774a3(0x151)+'i';}if(speed===_0x5774a3(0x163)&&quality==='good'){if(env[_0x5774a3(0x173)+'I_API'+_0x5774a3(0x16b)])return log('[Mode'+_0x5774a3(0x17a)+'Selec'+_0x5774a3(0x143)+_0x5774a3(0x141)+'+\x20goo'+_0x5774a3(0x16a)+_0x5774a3(0x15c)+_0x5774a3(0x174)+_0x5774a3(0x188)+_0x5774a3(0x148)+'ni'),_0x5774a3(0x176)+'i';}if(costSensitive){if(env['GEMIN'+_0x5774a3(0x15f)+'_KEY'])return log('[Mode'+_0x5774a3(0x17a)+'Selec'+'tor]\x20'+_0x5774a3(0x17b)+_0x5774a3(0x16f)+'tive,'+'\x20sele'+'cting'+_0x5774a3(0x148)+'ni'),_0x5774a3(0x176)+'i';if(env['GROQ_'+_0x5774a3(0x179)+'EY']&&!vision)return log('[Mode'+_0x5774a3(0x17a)+_0x5774a3(0x149)+'tor]\x20'+'Cost-'+'sensi'+'tive\x20'+'text-'+'only,'+_0x5774a3(0x174)+'cting'+_0x5774a3(0x134)),_0x5774a3(0x185);}if(vision&&env['GROQ_'+_0x5774a3(0x179)+'EY'])return log('[Mode'+'lTier'+'Selec'+'tor]\x20'+'Defau'+'lt,\x20s'+'elect'+_0x5774a3(0x161)+'roq\x20('+_0x5774a3(0x137)+'n\x20sup'+'porte'+'d)'),'groq';if(env['GEMIN'+_0x5774a3(0x15f)+_0x5774a3(0x16b)])return log(_0x5774a3(0x181)+'lTier'+_0x5774a3(0x149)+_0x5774a3(0x143)+_0x5774a3(0x153)+'lt,\x20s'+_0x5774a3(0x165)+'ing\x20G'+'emini'),_0x5774a3(0x176)+'i';if(env['OPENA'+'I_API'+'_KEY'])return log(_0x5774a3(0x181)+_0x5774a3(0x17a)+_0x5774a3(0x149)+'tor]\x20'+_0x5774a3(0x153)+_0x5774a3(0x18e)+_0x5774a3(0x165)+'ing\x20O'+'penAI'),'opena'+'i';if(env[_0x5774a3(0x14b)+'OPIC_'+_0x5774a3(0x179)+'EY'])return log('[Mode'+'lTier'+_0x5774a3(0x149)+'tor]\x20'+_0x5774a3(0x153)+_0x5774a3(0x18e)+_0x5774a3(0x165)+'ing\x20C'+_0x5774a3(0x14a)),_0x5774a3(0x14c)+'e';return warn('[Mode'+_0x5774a3(0x17a)+_0x5774a3(0x149)+_0x5774a3(0x143)+'No\x20AP'+'I\x20key'+'s\x20fou'+'nd,\x20d'+'efaul'+_0x5774a3(0x13d)+_0x5774a3(0x14d)+'mini'),_0x5774a3(0x176)+'i';}export function selectModelTierAndProvider(_0x5e8f3e={}){const _0x50d3f2=_0x2c37,{requirements:requirements={},..._0x31e560}=_0x5e8f3e,_0x5e580f=selectModelTier(_0x31e560),_0x5178d5={...requirements};_0x5178d5['env']=process[_0x50d3f2(0x14e)];const _0x2795ce=selectProvider(_0x5178d5),_0xef9135={};return _0xef9135['tier']=_0x5e580f,_0xef9135['provi'+'der']=_0x2795ce,_0xef9135['reaso'+'n']='Selec'+_0x50d3f2(0x152)+_0x2795ce+'\x20'+_0x5e580f+('\x20tier'+_0x50d3f2(0x136)+'d\x20on\x20'+'conte'+'xt'),_0xef9135;}

package/src/rubrics.mjs

@@ -72,12 +72,19 @@ export const DEFAULT_RUBRIC = {
 
 /**
  * Build rubric prompt section
- * 
+ *
  * @param {import('./index.mjs').Rubric | null} [rubric=null] - Rubric to use, or null for default
  * @param {boolean} [includeDimensions=true] - Whether to include evaluation dimensions
+ * @param {{ referenceImages?: Record<number, string> }} [options={}] - Options
+ *   referenceImages: map of score level -> image path for visual anchoring.
+ *   When provided, the rubric prompt instructs the VLM to compare against
+ *   reference images for each score level (Prometheus-Vision, arXiv:2401.06591).
+ *   The caller is responsible for encoding and attaching images to the API call;
+ *   this function only generates the text prompt referencing them.
  * @returns {string} Formatted rubric prompt text
  */
-export function buildRubricPrompt(rubric = null, includeDimensions = true) {
+export function buildRubricPrompt(rubric = null, includeDimensions = true, options = {}) {
+  const { referenceImages = null } = options;
   const rubricToUse = rubric || DEFAULT_RUBRIC;
   let prompt = `## EVALUATION RUBRIC
 
@@ -114,6 +121,19 @@ JSON: {"score": 3, "assessment": "fail", "issues": ["broken layout", "critical c
 7. List specific issues found (if any)
 8. Provide reasoning for your score`;
 
+  // Visual anchoring: reference images for score levels (Prometheus-Vision, arXiv:2401.06591)
+  if (referenceImages && typeof referenceImages === 'object') {
+    const levels = Object.keys(referenceImages).map(Number).sort((a, b) => b - a);
+    if (levels.length > 0) {
+      prompt += `\n\n### Visual Reference Anchors:
+The following reference images are provided as calibration anchors for specific score levels.
+Compare the screenshot being evaluated against these references to calibrate your scoring.
+${levels.map(level => `- **Score ${level}**: See reference image labeled "REF_SCORE_${level}"`).join('\n')}
+
+Use these references to anchor your absolute scores. A screenshot similar in quality to REF_SCORE_9 should score around 9, etc.`;
+    }
+  }
+
   if (includeDimensions && rubricToUse.dimensions) {
     prompt += `\n\n### Evaluation Dimensions:
 ${Object.entries(rubricToUse.dimensions)

package/src/score-calibration.mjs

@@ -0,0 +1,177 @@
+/**
+ * Score Calibration
+ *
+ * Adjusts raw VLM scores to reduce provider-specific bias.
+ * Research shows each VLM has a stable "evaluative fingerprint" --
+ * systematic scoring tendencies that differ across providers
+ * (Evaluative Fingerprints, arXiv:2601.05114).
+ *
+ * Supports:
+ * - Per-provider linear calibration (offset + scale)
+ * - User-supplied calibration profiles
+ * - Score histogram analysis for drift detection
+ */
+
+import { warn } from './logger.mjs';
+import { ValidationError } from './errors.mjs';
+
+/**
+ * Default calibration profiles per provider.
+ *
+ * These are initial estimates based on observed tendencies.
+ * Users should override with their own profiles via calibrate()
+ * after running createCalibrationSuite().
+ *
+ * Format: { offset, scale } where calibrated = (raw + offset) * scale
+ * Then clamped to [0, 10].
+ */
+const DEFAULT_PROFILES = {
+  gemini:    { offset: 0, scale: 1.0 },
+  openai:    { offset: 0, scale: 1.0 },
+  claude:    { offset: 0, scale: 1.0 },
+  groq:      { offset: 0, scale: 1.0 },
+  openrouter: { offset: 0, scale: 1.0 }
+};
+
+// User-supplied profiles override defaults
+let userProfiles = {};
+
+/**
+ * Set calibration profile for a provider
+ *
+ * @param {string} provider - Provider name
+ * @param {{ offset: number, scale: number }} profile - Calibration profile
+ */
+export function setCalibrationProfile(provider, profile) {
+  if (typeof profile.offset !== 'number' || typeof profile.scale !== 'number') {
+    throw new ValidationError('Calibration profile must have numeric offset and scale', { offset: typeof profile.offset, scale: typeof profile.scale });
+  }
+  if (profile.scale <= 0) {
+    throw new ValidationError('Calibration scale must be positive', { scale: profile.scale });
+  }
+  userProfiles[provider] = { ...profile };
+}
+
+/**
+ * Get calibration profile for a provider
+ *
+ * @param {string} provider - Provider name
+ * @returns {{ offset: number, scale: number }} Calibration profile
+ */
+export function getCalibrationProfile(provider) {
+  return userProfiles[provider] || DEFAULT_PROFILES[provider] || { offset: 0, scale: 1.0 };
+}
+
+/**
+ * Reset all calibration profiles to defaults
+ */
+export function resetCalibrationProfiles() {
+  userProfiles = {};
+}
+
+/**
+ * Calibrate a raw score using the provider's profile
+ *
+ * @param {number | null} score - Raw score from VLM (0-10)
+ * @param {string} provider - Provider name
+ * @returns {number | null} Calibrated score (0-10), or null if input is null
+ */
+export function calibrateScore(score, provider) {
+  if (score === null || score === undefined) {
+    return null;
+  }
+
+  const profile = getCalibrationProfile(provider);
+  const calibrated = (score + profile.offset) * profile.scale;
+
+  // Clamp to [0, 10]
+  return Math.max(0, Math.min(10, Math.round(calibrated * 100) / 100));
+}
+
+/**
+ * Derive a calibration profile from labeled data
+ *
+ * Given pairs of (raw VLM score, expected score), computes the
+ * least-squares linear fit: expected = raw * scale + offset.
+ *
+ * @param {Array<{ raw: number, expected: number }>} pairs - Score pairs
+ * @returns {{ offset: number, scale: number, r2: number }} Calibration profile with fit quality
+ */
+export function deriveCalibrationProfile(pairs) {
+  if (!Array.isArray(pairs) || pairs.length < 2) {
+    throw new ValidationError('Need at least 2 (raw, expected) pairs to derive calibration', { count: pairs?.length ?? 0 });
+  }
+
+  const n = pairs.length;
+  let sumX = 0, sumY = 0, sumXX = 0, sumXY = 0, sumYY = 0;
+
+  for (const { raw, expected } of pairs) {
+    sumX += raw;
+    sumY += expected;
+    sumXX += raw * raw;
+    sumXY += raw * expected;
+    sumYY += expected * expected;
+  }
+
+  const denom = n * sumXX - sumX * sumX;
+
+  if (Math.abs(denom) < 1e-10) {
+    warn('[Calibration] All raw scores are identical; cannot derive profile');
+    return { offset: 0, scale: 1.0, r2: 0 };
+  }
+
+  // Linear regression: expected = scale * raw + offset_intercept
+  // We want calibrated = (raw + offset) * scale, so:
+  // calibrated = raw * scale + offset * scale
+  // Matching: scale = slope, offset_intercept = offset * scale -> offset = intercept / scale
+  const slope = (n * sumXY - sumX * sumY) / denom;
+  const intercept = (sumY - slope * sumX) / n;
+
+  // Convert to our format: calibrated = (raw + offset) * scale
+  const scale = slope || 1.0;
+  const offset = scale !== 0 ? intercept / scale : 0;
+
+  // R-squared
+  const meanY = sumY / n;
+  const ssTot = sumYY - n * meanY * meanY;
+  const ssRes = pairs.reduce((sum, { raw, expected }) => {
+    const predicted = raw * slope + intercept;
+    return sum + (expected - predicted) ** 2;
+  }, 0);
+  const r2 = ssTot > 0 ? 1 - ssRes / ssTot : 0;
+
+  return { offset, scale, r2 };
+}
+
+/**
+ * Analyze score distribution for a provider to detect drift
+ *
+ * @param {number[]} scores - Array of scores from a single provider
+ * @returns {{ mean: number, stddev: number, skew: number, histogram: Record<number, number> }}
+ */
+export function analyzeScoreDistribution(scores) {
+  if (!scores.length) {
+    return { mean: 0, stddev: 0, skew: 0, histogram: {} };
+  }
+
+  const n = scores.length;
+  const mean = scores.reduce((a, b) => a + b, 0) / n;
+
+  const variance = scores.reduce((sum, s) => sum + (s - mean) ** 2, 0) / n;
+  const stddev = Math.sqrt(variance);
+
+  // Skewness (Fisher's)
+  const skew = stddev > 0
+    ? scores.reduce((sum, s) => sum + ((s - mean) / stddev) ** 3, 0) / n
+    : 0;
+
+  // Histogram (integer buckets 0-10)
+  const histogram = {};
+  for (let i = 0; i <= 10; i++) histogram[i] = 0;
+  for (const s of scores) {
+    const bucket = Math.max(0, Math.min(10, Math.round(s)));
+    histogram[bucket]++;
+  }
+
+  return { mean, stddev, skew, histogram };
+}