@arclabs561/ai-visual-test 0.7.4 → 0.7.5

package/CHANGELOG.md

@@ -2,6 +2,13 @@
 
 All notable changes to ai-visual-test will be documented in this file.
 
+## [0.7.5] - 2026-03-16
+
+### Fixed
+- **Test image generator**: replaced 1x1 pixel PNG stub with programmatic 100x100 gradient PNG using raw PNG chunk construction (zlib deflate). Fixes Groq API rejections ("Image must have at least 2 pixels") in integration tests.
+- **BatchOptimizer queue-full test**: matched error type to actual `ValidationError` thrown by `_queueRequest` (was expecting `TimeoutError`).
+- **Deleted stale test**: removed `validation-framework.test.mjs` referencing deleted `src/validation-framework.mjs` module.
+
 ## [0.7.4] - 2026-03-03
 
 ### Added

package/README.md

@@ -1,5 +1,8 @@
 # ai-visual-test
 
+[![npm](https://img.shields.io/npm/v/@arclabs561/ai-visual-test)](https://www.npmjs.com/package/@arclabs561/ai-visual-test)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
+
 Visual testing framework using Vision Language Models. Validates screenshots, checks accessibility, and can play games.
 
 ## Why This Package

package/index.d.ts

@@ -2116,3 +2116,143 @@ export function validateWithProgrammaticContext(
   options?: ValidationContext
 ): Promise<HybridValidationResult>;
 
+// --- Score Calibration (arXiv:2601.05114) ---
+
+export interface CalibrationProfile {
+  offset: number;
+  scale: number;
+}
+
+export interface DerivedCalibrationProfile extends CalibrationProfile {
+  r2: number;
+}
+
+export interface ScoreDistribution {
+  mean: number;
+  stddev: number;
+  skew: number;
+  histogram: Record<number, number>;
+}
+
+export function calibrateScore(score: number | null, provider: string): number | null;
+export function setCalibrationProfile(provider: string, profile: CalibrationProfile): void;
+export function getCalibrationProfile(provider: string): CalibrationProfile;
+export function resetCalibrationProfiles(): void;
+export function deriveCalibrationProfile(pairs: Array<{ raw: number; expected: number }>): DerivedCalibrationProfile;
+export function analyzeScoreDistribution(scores: number[]): ScoreDistribution;
+
+// --- Calibration Suite / Meta-evaluation (arXiv:2507.10062) ---
+
+export interface CalibrationSample {
+  screenshot: string;
+  expectedScore: number;
+  label?: string;
+  prompt?: string;
+}
+
+export interface CalibrationReport {
+  sampleCount: number;
+  scoredCount: number;
+  pearsonR: number | null;
+  spearmanR: number | null;
+  meanAbsoluteError: number | null;
+  maxError: number | null;
+  falsePositiveRate: number | null;
+  falseNegativeRate: number | null;
+  suggestedCalibration: DerivedCalibrationProfile | null;
+  scoreDistribution: ScoreDistribution;
+  details: Array<{
+    screenshot: string;
+    label: string;
+    expected: number;
+    actual: number | null;
+    error: number | null;
+    skipped?: string;
+  }>;
+  error?: string;
+}
+
+export interface CalibrationSuiteInstance {
+  samples: CalibrationSample[];
+  run(judgeOptions?: ConfigOptions): Promise<CalibrationReport>;
+}
+
+export function createCalibrationSuite(
+  samples: CalibrationSample[],
+  options?: { passThreshold?: number }
+): CalibrationSuiteInstance;
+
+// --- Known VLM Limitations (arXiv:2501.09236, arXiv:2511.03471) ---
+
+export interface VLMLimitation {
+  description: string;
+  severity: 'high' | 'medium' | 'low';
+  recommendation: string;
+  vlmAccuracy: 'none' | 'low' | 'medium' | 'high';
+}
+
+export type VLMLimitationKey =
+  | 'subtleSpatialShifts'
+  | 'elementOverlap'
+  | 'keyboardNavigation'
+  | 'screenReaderOrder'
+  | 'colorContrastPrecision'
+  | 'dynamicContent'
+  | 'textContent'
+  | 'interactiveState';
+
+export type TestType = 'accessibility' | 'layout' | 'visual' | 'interaction' | 'general';
+
+export const VLM_LIMITATIONS: Record<VLMLimitationKey, VLMLimitation>;
+
+export function getLimitationsForTestType(
+  testType: TestType
+): Array<{ key: string } & VLMLimitation>;
+
+export function shouldUseHybridValidation(testType: TestType): boolean;
+
+// --- Human Validation Manager ---
+
+export interface HumanValidationManagerOptions {
+  enabled?: boolean;
+  autoCollect?: boolean;
+  smartSampling?: boolean;
+  calibrationThreshold?: number;
+  humanValidatorFn?: ((vllmResult: ValidationResult) => Promise<unknown>) | null;
+}
+
+export class HumanValidationManager {
+  constructor(options?: HumanValidationManagerOptions);
+  enabled: boolean;
+  collectJudgment(result: ValidationResult): void;
+  requestValidation(result: ValidationResult): Promise<unknown>;
+  getCalibrationData(): Promise<unknown>;
+  isCalibrated(): boolean;
+}
+
+export function getHumanValidationManager(): HumanValidationManager;
+export function initHumanValidation(options?: HumanValidationManagerOptions): HumanValidationManager;
+
+// --- Explanation Manager ---
+
+export class ExplanationManager {
+  constructor(options?: ConfigOptions);
+  explainJudgment(vllmJudgment: object, question?: string | null, options?: object): Promise<object>;
+}
+
+export function getExplanationManager(options?: ConfigOptions): ExplanationManager;
+
+// --- Temporal Batch Optimizer ---
+
+export class TemporalBatchOptimizer extends BatchOptimizer {
+  constructor(options?: object);
+  addWithDependencies(request: object, dependencies?: string[]): Promise<ValidationResult>;
+}
+
+// --- Latency-Aware Batch Optimizer ---
+
+export class LatencyAwareBatchOptimizer extends BatchOptimizer {
+  constructor(options?: { maxConcurrency?: number; batchSize?: number; cacheEnabled?: boolean; defaultMaxLatency?: number; adaptiveBatchSize?: boolean });
+  addWithLatencyTarget(request: object, maxLatencyMs?: number): Promise<ValidationResult>;
+}
+

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@arclabs561/ai-visual-test",
-  "version": "0.7.4",
+  "version": "0.7.5",
   "description": "Visual testing framework for web applications using Vision Language Models",
   "type": "module",
   "main": "src/index.mjs",
@@ -39,11 +39,7 @@
   "author": "arclabs561 <henry@henrywallace.io>",
   "license": "MIT",
   "dependencies": {
-    "@anthropic-ai/sdk": "0.70.0",
-    "@google/generative-ai": "0.24.1",
-    "async-mutex": "0.5.0",
-    "dotenv": "^16.4.5",
-    "openai": "6.9.1"
+    "async-mutex": "0.5.0"
   },
   "peerDependencies": {
     "@arclabs561/llm-utils": "*",

package/src/batch-optimizer.mjs

@@ -29,7 +29,7 @@ import { createHash } from 'crypto';
  * @class BatchOptimizer
  */
 import { API_CONSTANTS, BATCH_OPTIMIZER_CONSTANTS } from './constants.mjs';
-import { TimeoutError } from './errors.mjs';
+import { TimeoutError, ValidationError } from './errors.mjs';
 import { warn } from './logger.mjs';
 
 export class BatchOptimizer {
@@ -103,7 +103,7 @@ export class BatchOptimizer {
     const keyData = {
       imagePath,
       prompt: prompt || '',
-      context: context ? JSON.stringify(context) : ''
+      context: context ? JSON.stringify(context, Object.keys(context).sort()) : ''
     };
     const keyString = JSON.stringify(keyData);
     return createHash('sha256').update(keyString).digest('hex');
@@ -238,7 +238,7 @@ export class BatchOptimizer {
         });
       
       warn(`[BatchOptimizer] Queue is full (${this.queue.length}/${this.maxQueueSize}). Rejecting request to prevent memory leak. Total rejections: ${this.metrics.queueRejections}`);
-      throw new TimeoutError(
+      throw new ValidationError(
         `Queue is full (${this.queue.length}/${this.maxQueueSize}). Too many concurrent requests.`,
         { queueSize: this.queue.length, maxQueueSize: this.maxQueueSize }
       );

package/src/cache.mjs

@@ -55,12 +55,11 @@ let cacheMetrics = { atomicWrites: 0, atomicWriteFailures: 0, tempFileCleanups:
 export function initCache(cacheDir) {
   // SECURITY: Validate and normalize cache directory to prevent path traversal
   if (cacheDir) {
-    const normalized = normalize(resolve(cacheDir));
-    // Prevent path traversal
-    if (normalized.includes('..')) {
+    // Prevent path traversal: reject raw input containing '..' segments
+    if (cacheDir.includes('..')) {
       throw new CacheError('Invalid cache directory: path traversal detected', { cacheDir });
     }
-    CACHE_DIR = normalized;
+    CACHE_DIR = normalize(resolve(cacheDir));
   } else {
     CACHE_DIR = join(__dirname, '..', '..', '..', 'test-results', 'vllm-cache');
   }

package/src/calibration-suite.mjs

@@ -0,0 +1,197 @@
+/**
+ * Calibration Suite (Meta-evaluation)
+ *
+ * Measures how well a VLM judge performs against human-labeled ground truth.
+ * When using ML as a test oracle, you need tests of the tests
+ * (LLMShot, arXiv:2507.10062; Stanford CS 329T).
+ *
+ * Usage:
+ *   const suite = createCalibrationSuite([
+ *     { screenshot: 'good.png', expectedScore: 9, label: 'clean homepage' },
+ *     { screenshot: 'broken.png', expectedScore: 2, label: 'broken layout' },
+ *   ]);
+ *   const report = await suite.run({ provider: 'gemini' });
+ *   // report.correlation, report.meanAbsoluteError, report.falsePositiveRate, ...
+ */
+
+import { VLLMJudge } from './judge.mjs';
+import { ValidationError, ConfigError } from './errors.mjs';
+import { warn } from './logger.mjs';
+import { pearsonCorrelation, spearmanCorrelation } from './metrics.mjs';
+import { deriveCalibrationProfile, analyzeScoreDistribution } from './score-calibration.mjs';
+
+/**
+ * @typedef {Object} CalibrationSample
+ * @property {string} screenshot - Path to screenshot file
+ * @property {number} expectedScore - Human-labeled expected score (0-10)
+ * @property {string} [label] - Description of the sample
+ * @property {string} [prompt] - Custom prompt (defaults to generic quality evaluation)
+ */
+
+/**
+ * @typedef {Object} CalibrationReport
+ * @property {number} sampleCount - Number of samples evaluated
+ * @property {number} pearsonR - Pearson correlation between VLM and expected scores
+ * @property {number} spearmanR - Spearman rank correlation
+ * @property {number} meanAbsoluteError - Average |VLM score - expected score|
+ * @property {number} maxError - Worst single-sample error
+ * @property {number} falsePositiveRate - Rate of VLM pass (>=7) when expected fail (<7)
+ * @property {number} falseNegativeRate - Rate of VLM fail (<7) when expected pass (>=7)
+ * @property {{ offset: number, scale: number, r2: number }} suggestedCalibration - Derived calibration profile
+ * @property {{ mean: number, stddev: number, skew: number }} scoreDistribution - VLM score distribution
+ * @property {Array<{ screenshot: string, label: string, expected: number, actual: number, error: number }>} details
+ */
+
+const DEFAULT_PROMPT = 'Evaluate the overall quality of this screenshot. Consider layout, visual design, readability, accessibility, and functional correctness.';
+const DEFAULT_PASS_THRESHOLD = 7;
+
+/**
+ * Create a calibration suite
+ *
+ * @param {CalibrationSample[]} samples - Labeled screenshot samples
+ * @param {{ passThreshold?: number }} [options={}] - Suite options
+ * @returns {{ run: (judgeOptions: object) => Promise<CalibrationReport>, samples: CalibrationSample[] }}
+ */
+export function createCalibrationSuite(samples, options = {}) {
+  if (!Array.isArray(samples) || samples.length < 2) {
+    throw new ValidationError('Calibration suite requires at least 2 labeled samples', { count: samples?.length ?? 0 });
+  }
+
+  for (const s of samples) {
+    if (typeof s.screenshot !== 'string' || typeof s.expectedScore !== 'number') {
+      throw new ValidationError('Each sample must have a screenshot path (string) and expectedScore (number)', { screenshot: typeof s.screenshot, expectedScore: typeof s.expectedScore });
+    }
+    if (s.expectedScore < 0 || s.expectedScore > 10) {
+      throw new ValidationError(`expectedScore must be 0-10, got ${s.expectedScore}`, { expectedScore: s.expectedScore });
+    }
+  }
+
+  const passThreshold = options.passThreshold ?? DEFAULT_PASS_THRESHOLD;
+
+  return {
+    samples: [...samples],
+
+    /**
+     * Run the calibration suite against a judge
+     *
+     * @param {object} judgeOptions - Options passed to VLLMJudge constructor
+     * @returns {Promise<CalibrationReport>}
+     */
+    async run(judgeOptions = {}) {
+      const judge = new VLLMJudge(judgeOptions);
+
+      if (!judge.enabled) {
+        throw new ConfigError('VLLMJudge is disabled -- cannot run calibration suite');
+      }
+
+      const details = [];
+      const rawScores = [];
+      const expectedScores = [];
+
+      for (const sample of samples) {
+        const prompt = sample.prompt || DEFAULT_PROMPT;
+        try {
+          const result = await judge.judgeScreenshot(sample.screenshot, prompt);
+          const parsed = judge.extractSemanticInfo(result.judgment || '');
+          const score = parsed.score ?? result.score ?? null;
+
+          if (score !== null) {
+            const error = Math.abs(score - sample.expectedScore);
+            rawScores.push(score);
+            expectedScores.push(sample.expectedScore);
+            details.push({
+              screenshot: sample.screenshot,
+              label: sample.label || sample.screenshot,
+              expected: sample.expectedScore,
+              actual: score,
+              error
+            });
+          } else {
+            details.push({
+              screenshot: sample.screenshot,
+              label: sample.label || sample.screenshot,
+              expected: sample.expectedScore,
+              actual: null,
+              error: null,
+              skipped: 'No score returned'
+            });
+          }
+        } catch (err) {
+          warn(`[CalibrationSuite] Sample "${sample.label || sample.screenshot}" failed: ${err.message}`);
+          details.push({
+            screenshot: sample.screenshot,
+            label: sample.label || sample.screenshot,
+            expected: sample.expectedScore,
+            actual: null,
+            error: null,
+            errorMessage: err.message,
+            skipped: true
+          });
+        }
+      }
+
+      // Compute metrics from scored samples only
+      const scored = details.filter(d => d.actual !== null);
+      const n = scored.length;
+
+      if (n < 2) {
+        return {
+          sampleCount: samples.length,
+          scoredCount: n,
+          pearsonR: null,
+          spearmanR: null,
+          meanAbsoluteError: null,
+          maxError: null,
+          falsePositiveRate: null,
+          falseNegativeRate: null,
+          suggestedCalibration: null,
+          scoreDistribution: analyzeScoreDistribution(rawScores),
+          details,
+          error: 'Fewer than 2 samples scored -- cannot compute metrics'
+        };
+      }
+
+      const errors = scored.map(d => d.error);
+      const meanAbsoluteError = errors.reduce((a, b) => a + b, 0) / n;
+      const maxError = Math.max(...errors);
+
+      // Correlation
+      const actualScores = scored.map(d => d.actual);
+      const expScores = scored.map(d => d.expected);
+      const pearsonR = pearsonCorrelation(actualScores, expScores);
+      const spearmanR = spearmanCorrelation(actualScores, expScores);
+
+      // False positive/negative rates (relative to passThreshold)
+      let fp = 0, fn = 0, expectedFail = 0, expectedPass = 0;
+      for (const d of scored) {
+        if (d.expected >= passThreshold) {
+          expectedPass++;
+          if (d.actual < passThreshold) fn++;
+        } else {
+          expectedFail++;
+          if (d.actual >= passThreshold) fp++;
+        }
+      }
+      const falsePositiveRate = expectedFail > 0 ? fp / expectedFail : 0;
+      const falseNegativeRate = expectedPass > 0 ? fn / expectedPass : 0;
+
+      // Derive calibration profile
+      const pairs = scored.map(d => ({ raw: d.actual, expected: d.expected }));
+      const suggestedCalibration = deriveCalibrationProfile(pairs);
+
+      return {
+        sampleCount: samples.length,
+        scoredCount: n,
+        pearsonR,
+        spearmanR,
+        meanAbsoluteError,
+        maxError,
+        falsePositiveRate,
+        falseNegativeRate,
+        suggestedCalibration,
+        scoreDistribution: analyzeScoreDistribution(rawScores),
+        details
+      };
+    }
+  };
+}

package/src/constants.mjs

@@ -98,6 +98,17 @@ export const API_ENDPOINT_CONSTANTS = {
   RATE_LIMIT_MAX_REQUESTS: 10
 };
 
+/**
+ * Rate Limit Validation Bounds
+ */
+export const RATE_LIMIT_BOUNDS = {
+  /** Minimum allowed value for RATE_LIMIT_MAX_REQUESTS */
+  MIN: 1,
+
+  /** Maximum allowed value for RATE_LIMIT_MAX_REQUESTS */
+  MAX: 1000
+};
+
 /**
  * Retry Configuration
  */

package/src/cost-optimization.mjs

@@ -1 +1 @@
-(function(_0x18385b,_0x2f2f0a){const _0xa53b50=_0x2d2e,_0x49ca27=_0x18385b();while(!![]){try{const _0xdd5ddf=parseInt(_0xa53b50(0x171))/0x1*(parseInt(_0xa53b50(0x16e))/0x2)+parseInt(_0xa53b50(0x199))/0x3*(-parseInt(_0xa53b50(0x174))/0x4)+parseInt(_0xa53b50(0x19c))/0x5*(parseInt(_0xa53b50(0x180))/0x6)+parseInt(_0xa53b50(0x16b))/0x7*(-parseInt(_0xa53b50(0x195))/0x8)+-parseInt(_0xa53b50(0x17c))/0x9*(-parseInt(_0xa53b50(0x197))/0xa)+-parseInt(_0xa53b50(0x18b))/0xb*(-parseInt(_0xa53b50(0x177))/0xc)+parseInt(_0xa53b50(0x17a))/0xd;if(_0xdd5ddf===_0x2f2f0a)break;else _0x49ca27['push'](_0x49ca27['shift']());}catch(_0x16dc43){_0x49ca27['push'](_0x49ca27['shift']());}}}(_0x1ecb,0x4479d));const _0x40f56c=(function(){let _0x180237=!![];return function(_0x539394,_0xb40faa){const _0x42db3d=_0x180237?function(){if(_0xb40faa){const _0x4b441b=_0xb40faa['apply'](_0x539394,arguments);return _0xb40faa=null,_0x4b441b;}}:function(){};return _0x180237=![],_0x42db3d;};}()),_0xd2e06c=_0x40f56c(this,function(){const _0x212118=_0x2d2e;return _0xd2e06c[_0x212118(0x18c)+'ing']()['searc'+'h'](_0x212118(0x193)+_0x212118(0x188)+'+$')['toStr'+_0x212118(0x198)]()['const'+'ructo'+'r'](_0xd2e06c)[_0x212118(0x167)+'h'](_0x212118(0x193)+_0x212118(0x188)+'+$');});_0xd2e06c();import{selectModelTier,selectProvider,selectModelTierAndProvider}from'./model-tier-selector.mjs';import{createConfig,getProvider}from'./config.mjs';export function calculateCostComparison(_0x57230a={},_0x52fd2f={}){const _0xaa590d=_0x2d2e,_0x33bd6a=parseFloat(_0x52fd2f[_0xaa590d(0x184)+'atedC'+_0xaa590d(0x168)]?.['total'+'Cost']||'0'),_0x397464=_0x57230a['model'+_0xaa590d(0x1a7)]||'balan'+_0xaa590d(0x173),_0x238e4b=_0x52fd2f['provi'+'der']||'gemin'+'i',_0x5ef67d=getProvider(_0x238e4b),_0x48530e={};_0x48530e['input']=0x0,_0x48530e['outpu'+'t']=0x0;const _0x26c913=_0x5ef67d?.[_0xaa590d(0x181)+'ng']||_0x48530e,_0x4f9c20=0x3e8,_0x321d8d=0x1f4,_0x578480={};for(const _0x1ff61d of[_0xaa590d(0x182),_0xaa590d(0x19b)+'ced','best']){const _0x161f86=_0x4f9c20/0xf4240*_0x26c913['input'],_0x519e8b=_0x321d8d/0xf4240*_0x26c913[_0xaa590d(0x1a1)+'t'];_0x578480[_0x1ff61d]=_0x161f86+_0x519e8b;}const _0x420aba={};for(const _0xf4fc80 of[_0xaa590d(0x182),'balan'+_0xaa590d(0x173),_0xaa590d(0x1a9)]){if(_0x578480[_0xf4fc80]&&_0x33bd6a>0x0){const _0x1573e6=_0x33bd6a-_0x578480[_0xf4fc80],_0x36dd72=_0x1573e6/_0x33bd6a*0x64;_0x420aba[_0xf4fc80]={'absolute':_0x1573e6,'percent':_0x36dd72,'cost':_0x578480[_0xf4fc80]};}}const _0x31be62={};return _0x31be62['tier']=_0x397464,_0x31be62[_0xaa590d(0x169)+'der']=_0x238e4b,_0x31be62[_0xaa590d(0x1a4)]=_0x33bd6a,{'current':_0x31be62,'tiers':_0x578480,'savings':_0x420aba,'recommendation':getCostOptimizationRecommendation(_0x57230a,_0x33bd6a,_0x578480)};}function getCostOptimizationRecommendation(_0x48b039,_0x19d13e,_0x2fc7c2){const _0x509a09=_0x2d2e,{frequency:_0x3a7aef,criticality:_0x1ccde8,costSensitive:_0x4f54d4}=_0x48b039;let _0x2bcee6=_0x509a09(0x19b)+_0x509a09(0x173);if(_0x3a7aef===_0x509a09(0x18e)||_0x3a7aef>=0xa||_0x4f54d4)_0x2bcee6=_0x509a09(0x182);else _0x1ccde8===_0x509a09(0x18f)+_0x509a09(0x194)&&(_0x2bcee6='best');const _0x505d15=_0x2fc7c2[_0x2bcee6]||_0x19d13e,_0x44afb3=_0x19d13e-_0x505d15,_0x10c8ba=_0x19d13e>0x0?_0x44afb3/_0x19d13e*0x64:0x0;return{'tier':_0x2bcee6,'cost':_0x505d15,'savings':_0x44afb3,'savingsPercent':_0x10c8ba,'reason':getRecommendationReason(_0x48b039,_0x2bcee6)};}function getRecommendationReason(_0x598c0e,_0xf227a9){const _0x30edec=_0x2d2e;if(_0xf227a9==='fast'){if(_0x598c0e[_0x30edec(0x185)+'ency']===_0x30edec(0x18e)||_0x598c0e['frequ'+_0x30edec(0x1a2)]>=0xa)return _0x30edec(0x1a5)+_0x30edec(0x185)+'ency\x20'+_0x30edec(0x17d)+_0x30edec(0x187)+'\x20requ'+_0x30edec(0x196)+_0x30edec(0x179)+_0x30edec(0x192);if(_0x598c0e[_0x30edec(0x189)+'ensit'+'ive'])return _0x30edec(0x1ac)+'sensi'+'tive\x20'+_0x30edec(0x183)+_0x30edec(0x1ab)+'\x20use\x20'+_0x30edec(0x179)+_0x30edec(0x192);}if(_0xf227a9==='best')return'Criti'+_0x30edec(0x178)+'valua'+_0x30edec(0x190)+_0x30edec(0x16f)+'res\x20b'+'est\x20t'+_0x30edec(0x1a3)+'or\x20qu'+_0x30edec(0x1aa);return'Balan'+_0x30edec(0x18a)+_0x30edec(0x170)+'rovid'+'es\x20sp'+'eed/q'+_0x30edec(0x1a6)+'y\x20tra'+'deoff';}function _0x2d2e(_0x2bf370,_0x100b7e){const _0x3462a3=_0x1ecb();return _0x2d2e=function(_0xd2e06c,_0x40f56c){_0xd2e06c=_0xd2e06c-0x167;let _0x1ecb23=_0x3462a3[_0xd2e06c];if(_0x2d2e['clmymu']===undefined){var _0x2d2e4c=function(_0x4d9406){const _0x4347ab='abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789+/=';let _0x91f6aa='',_0x4491a1='',_0x428b38=_0x91f6aa+_0x2d2e4c;for(let _0x2de447=0x0,_0x464545,_0x5ec257,_0x372eca=0x0;_0x5ec257=_0x4d9406['charAt'](_0x372eca++);~_0x5ec257&&(_0x464545=_0x2de447%0x4?_0x464545*0x40+_0x5ec257:_0x5ec257,_0x2de447++%0x4)?_0x91f6aa+=_0x428b38['charCodeAt'](_0x372eca+0xa)-0xa!==0x0?String['fromCharCode'](0xff&_0x464545>>(-0x2*_0x2de447&0x6)):_0x2de447:0x0){_0x5ec257=_0x4347ab['indexOf'](_0x5ec257);}for(let _0x180237=0x0,_0x539394=_0x91f6aa['length'];_0x180237<_0x539394;_0x180237++){_0x4491a1+='%'+('00'+_0x91f6aa['charCodeAt'](_0x180237)['toString'](0x10))['slice'](-0x2);}return decodeURIComponent(_0x4491a1);};_0x2d2e['YBBLwq']=_0x2d2e4c,_0x2bf370=arguments,_0x2d2e['clmymu']=!![];}const _0x5bd10d=_0x3462a3[0x0],_0x25f117=_0xd2e06c+_0x5bd10d,_0x4bfa19=_0x2bf370[_0x25f117];if(!_0x4bfa19){const _0xb40faa=function(_0x42db3d){this['tUiOdk']=_0x42db3d,this['yllaQc']=[0x1,0x0,0x0],this['eWSZJt']=function(){return'newState';},this['NeSNJo']='\x5cw+\x20*\x5c(\x5c)\x20*{\x5cw+\x20*',this['jVFOWK']='[\x27|\x22].+[\x27|\x22];?\x20*}';};_0xb40faa['prototype']['JTaKsI']=function(){const _0x4b441b=new RegExp(this['NeSNJo']+this['jVFOWK']),_0x57230a=_0x4b441b['test'](this['eWSZJt']['toString']())?--this['yllaQc'][0x1]:--this['yllaQc'][0x0];return this['eHArMZ'](_0x57230a);},_0xb40faa['prototype']['eHArMZ']=function(_0x52fd2f){if(!Boolean(~_0x52fd2f))return _0x52fd2f;return this['AIYzCx'](this['tUiOdk']);},_0xb40faa['prototype']['AIYzCx']=function(_0x33bd6a){for(let _0x397464=0x0,_0x238e4b=this['yllaQc']['length'];_0x397464<_0x238e4b;_0x397464++){this['yllaQc']['push'](Math['round'](Math['random']())),_0x238e4b=this['yllaQc']['length'];}return _0x33bd6a(this['yllaQc'][0x0]);},new _0xb40faa(_0x2d2e)['JTaKsI'](),_0x1ecb23=_0x2d2e['YBBLwq'](_0x1ecb23),_0x2bf370[_0x25f117]=_0x1ecb23;}else _0x1ecb23=_0x4bfa19;return _0x1ecb23;},_0x2d2e(_0x2bf370,_0x100b7e);}export function optimizeCost(_0x157b14={}){const _0x46d2fd=_0x2d2e,{frequency:_0x11c64d,criticality:_0x1162ee,costSensitive:_0x2e4e15,budget:_0x2fbb39,requirements:requirements={}}=_0x157b14,_0x5d2c5d={};_0x5d2c5d[_0x46d2fd(0x185)+'ency']=_0x11c64d,_0x5d2c5d[_0x46d2fd(0x18f)+_0x46d2fd(0x1a0)+'y']=_0x1162ee,_0x5d2c5d['costS'+'ensit'+'ive']=_0x2e4e15,_0x5d2c5d[_0x46d2fd(0x16f)+_0x46d2fd(0x186)+'ts']={...requirements},_0x5d2c5d[_0x46d2fd(0x16f)+_0x46d2fd(0x186)+'ts'][_0x46d2fd(0x189)+'ensit'+'ive']=_0x2e4e15,_0x5d2c5d[_0x46d2fd(0x16f)+_0x46d2fd(0x186)+'ts']['env']=process[_0x46d2fd(0x17f)];const {tier:_0x214f9b,provider:_0x20c49a,reason:_0x5c1bce}=selectModelTierAndProvider(_0x5d2c5d),_0xdaa5b8={};_0xdaa5b8['model'+'Tier']=_0x214f9b,_0xdaa5b8[_0x46d2fd(0x169)+_0x46d2fd(0x19a)]=_0x20c49a;const _0xfb6ace=createConfig(_0xdaa5b8),_0x56ddaf=getProvider(_0x20c49a),_0x409eff={};_0x409eff['input']=0x0,_0x409eff['outpu'+'t']=0x0;const _0x422e41=_0x56ddaf?.[_0x46d2fd(0x181)+'ng']||_0x409eff,_0x117a81=0x3e8,_0x510ec9=0x1f4,_0x59906d=_0x117a81/0xf4240*_0x422e41[_0x46d2fd(0x16c)]+_0x510ec9/0xf4240*_0x422e41[_0x46d2fd(0x1a1)+'t'],_0x39264a={};for(const _0x149af7 of['fast',_0x46d2fd(0x19b)+'ced','best']){if(_0x149af7!==_0x214f9b){const _0x87f361=_0x59906d,_0x3fad65={};_0x3fad65['cost']=_0x87f361,_0x3fad65[_0x46d2fd(0x19e)+'gs']=_0x59906d-_0x87f361,_0x3fad65[_0x46d2fd(0x19e)+'gsPer'+_0x46d2fd(0x19f)]=_0x59906d>0x0?(_0x59906d-_0x87f361)/_0x59906d*0x64:0x0,_0x39264a[_0x149af7]=_0x3fad65;}}const _0x449d7a=_0x2fbb39?_0x59906d<=_0x2fbb39:null;return{'recommendedTier':_0x214f9b,'recommendedProvider':_0x20c49a,'estimatedCost':_0x59906d,'savings':getSavingsEstimate(_0x214f9b,_0x20c49a,_0x39264a),'config':_0xfb6ace,'reason':_0x5c1bce,'withinBudget':_0x449d7a,'comparisons':_0x39264a,'recommendation':_0x449d7a===![]?_0x46d2fd(0x191)+'ated\x20'+'cost\x20'+'($'+_0x59906d[_0x46d2fd(0x1ad)+'ed'](0x6)+(')\x20exc'+_0x46d2fd(0x175)+'budge'+'t\x20($')+_0x2fbb39[_0x46d2fd(0x1ad)+'ed'](0x6)+(_0x46d2fd(0x17b)+_0x46d2fd(0x172)+'r\x20usi'+'ng\x20\x27f'+_0x46d2fd(0x16d)+'tier.'):'Optim'+'al\x20co'+_0x46d2fd(0x18d)+'ratio'+'n:\x20'+_0x20c49a+'\x20'+_0x214f9b+(_0x46d2fd(0x19d)+'\x20(est'+'imate'+_0x46d2fd(0x1ae))+_0x59906d[_0x46d2fd(0x1ad)+'ed'](0x6)+('\x20per\x20'+'valid'+_0x46d2fd(0x187)+')')};}function getSavingsEstimate(_0x1521bc,_0x3e64cf,_0x151d7f){const _0x373a2c=_0x2d2e;if(_0x1521bc===_0x373a2c(0x182)){const _0x36077e=_0x151d7f['balan'+_0x373a2c(0x173)]?.['savin'+'gs']||0x0,_0xbf31ee=_0x151d7f['best']?.[_0x373a2c(0x19e)+'gs']||0x0;return{'vsBalanced':_0x36077e>0x0?(_0x151d7f[_0x373a2c(0x19b)+_0x373a2c(0x173)]['savin'+_0x373a2c(0x16a)+_0x373a2c(0x19f)]||0x0)['toFix'+'ed'](0x0)+'%':'0%','vsBest':_0xbf31ee>0x0?(_0x151d7f[_0x373a2c(0x1a9)][_0x373a2c(0x19e)+_0x373a2c(0x16a)+_0x373a2c(0x19f)]||0x0)[_0x373a2c(0x1ad)+'ed'](0x0)+'%':'0%'};}if(_0x1521bc===_0x373a2c(0x19b)+'ced'){const _0xa4c8f8=_0x151d7f['fast']?.['savin'+'gs']||0x0,_0x23d59f=_0x151d7f['best']?.[_0x373a2c(0x19e)+'gs']||0x0;return{'vsFast':_0xa4c8f8<0x0?Math[_0x373a2c(0x1a8)](_0x151d7f[_0x373a2c(0x182)]?.['savin'+_0x373a2c(0x16a)+_0x373a2c(0x19f)]||0x0)['toFix'+'ed'](0x0)+(_0x373a2c(0x17e)+'e\x20exp'+_0x373a2c(0x176)+'e'):'0%','vsBest':_0x23d59f>0x0?(_0x151d7f['best'][_0x373a2c(0x19e)+'gsPer'+'cent']||0x0)[_0x373a2c(0x1ad)+'ed'](0x0)+'%':'0%'};}return{'vsFast':_0x151d7f[_0x373a2c(0x182)]?Math['abs'](_0x151d7f[_0x373a2c(0x182)]['savin'+_0x373a2c(0x16a)+'cent']||0x0)['toFix'+'ed'](0x0)+(_0x373a2c(0x17e)+'e\x20exp'+'ensiv'+'e'):'0%','vsBalanced':_0x151d7f[_0x373a2c(0x19b)+'ced']?Math['abs'](_0x151d7f[_0x373a2c(0x19b)+'ced'][_0x373a2c(0x19e)+_0x373a2c(0x16a)+'cent']||0x0)['toFix'+'ed'](0x0)+('%\x20mor'+'e\x20exp'+'ensiv'+'e'):'0%'};}function _0x1ecb(){const _0xa9d53f=['B3v0Chu','zw5JEq','AwvYigy','y29ZDa','sgLNAc0','DwfSAxq','vgLLCG','ywjZ','yMvZDa','ywXPDhK','DgLVBIW','q29ZDc0','Dg9gAxG','zdOGja','C2vHCMm','B3n0','ChjVDMK','z3nqzxi','mti2zgnotfDj','Aw5WDxq','yxn0jYa','nJyYogXUA1nRBq','CMvXDwK','AwvYiha','mtjgCNb5Cxe','BNnPzgu','y2vK','mtaYmtzJB3bkChu','zwvKCYa','zw5ZAxy','mtq2nZzxvgnduMO','y2fSigu','zMfZDca','ntKWmZnwsxPhruO','ks4Gq28','ndqWnZnkBhHUsK0','DMfSAwq','jsbTB3i','zw52','mZC5ogrsuuvgza','ChjPy2K','zMfZDa','B3bLCMe','zxn0Aw0','zNjLCxu','CMvTzw4','yxrPB24','ksSPkYK','y29ZDfm','y2vKihq','ndK5nezSrvjpDW','Dg9tDhi','BMzPz3u','AgLNAa','y3jPDgK','DgLVBIa','rxn0Aw0','DgLLCG','kcGOlIS','y2fS','mJiYntq0r0vtywDL','AxjLCYa','odKWCfPWyNfX','Aw5N','mZKZBhLSD3jo','zgvY','yMfSyw4','nJm1qwHeEeTT','ihrPzxi','C2f2Aw4','y2vUDa','y2fSAxq'];_0x1ecb=function(){return _0xa9d53f;};return _0x1ecb();}
+function _0x44fa(_0x1b7e6c,_0x4f2e91){const _0x39b127=_0x41cc();return _0x44fa=function(_0x42fad3,_0x7d21ab){_0x42fad3=_0x42fad3-0xa0;let _0x41cca5=_0x39b127[_0x42fad3];if(_0x44fa['mZqzuW']===undefined){var _0x44fa51=function(_0x137d0a){const _0x9cebc5='abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789+/=';let _0x10951a='',_0xb545cc='',_0x3eaa48=_0x10951a+_0x44fa51;for(let _0x5ce45a=0x0,_0x4393b2,_0x28fb3f,_0x197a35=0x0;_0x28fb3f=_0x137d0a['charAt'](_0x197a35++);~_0x28fb3f&&(_0x4393b2=_0x5ce45a%0x4?_0x4393b2*0x40+_0x28fb3f:_0x28fb3f,_0x5ce45a++%0x4)?_0x10951a+=_0x3eaa48['charCodeAt'](_0x197a35+0xa)-0xa!==0x0?String['fromCharCode'](0xff&_0x4393b2>>(-0x2*_0x5ce45a&0x6)):_0x5ce45a:0x0){_0x28fb3f=_0x9cebc5['indexOf'](_0x28fb3f);}for(let _0x3eae9f=0x0,_0x50f621=_0x10951a['length'];_0x3eae9f<_0x50f621;_0x3eae9f++){_0xb545cc+='%'+('00'+_0x10951a['charCodeAt'](_0x3eae9f)['toString'](0x10))['slice'](-0x2);}return decodeURIComponent(_0xb545cc);};_0x44fa['ctDYOL']=_0x44fa51,_0x1b7e6c=arguments,_0x44fa['mZqzuW']=!![];}const _0x91b6e8=_0x39b127[0x0],_0x5a6c85=_0x42fad3+_0x91b6e8,_0x1549a6=_0x1b7e6c[_0x5a6c85];if(!_0x1549a6){const _0x4de8b5=function(_0x193daf){this['bmzfVw']=_0x193daf,this['pBTXLc']=[0x1,0x0,0x0],this['loDRVx']=function(){return'newState';},this['dKehkO']='\x5cw+\x20*\x5c(\x5c)\x20*{\x5cw+\x20*',this['vrQtQR']='[\x27|\x22].+[\x27|\x22];?\x20*}';};_0x4de8b5['prototype']['ZpGEKi']=function(){const _0x504a68=new RegExp(this['dKehkO']+this['vrQtQR']),_0xa11c38=_0x504a68['test'](this['loDRVx']['toString']())?--this['pBTXLc'][0x1]:--this['pBTXLc'][0x0];return this['WIOdvz'](_0xa11c38);},_0x4de8b5['prototype']['WIOdvz']=function(_0x37beac){if(!Boolean(~_0x37beac))return _0x37beac;return this['OPbltl'](this['bmzfVw']);},_0x4de8b5['prototype']['OPbltl']=function(_0x4cd724){for(let _0x352e1c=0x0,_0x31d3e2=this['pBTXLc']['length'];_0x352e1c<_0x31d3e2;_0x352e1c++){this['pBTXLc']['push'](Math['round'](Math['random']())),_0x31d3e2=this['pBTXLc']['length'];}return _0x4cd724(this['pBTXLc'][0x0]);},new _0x4de8b5(_0x44fa)['ZpGEKi'](),_0x41cca5=_0x44fa['ctDYOL'](_0x41cca5),_0x1b7e6c[_0x5a6c85]=_0x41cca5;}else _0x41cca5=_0x1549a6;return _0x41cca5;},_0x44fa(_0x1b7e6c,_0x4f2e91);}(function(_0x440952,_0x534470){const _0x5c6aa1=_0x44fa,_0x19edde=_0x440952();while(!![]){try{const _0xbf939d=-parseInt(_0x5c6aa1(0xb5))/0x1*(parseInt(_0x5c6aa1(0xc5))/0x2)+parseInt(_0x5c6aa1(0xd9))/0x3*(parseInt(_0x5c6aa1(0xb6))/0x4)+-parseInt(_0x5c6aa1(0xb1))/0x5*(parseInt(_0x5c6aa1(0xa2))/0x6)+-parseInt(_0x5c6aa1(0xa3))/0x7+parseInt(_0x5c6aa1(0xc6))/0x8+-parseInt(_0x5c6aa1(0xe0))/0x9+parseInt(_0x5c6aa1(0xbc))/0xa;if(_0xbf939d===_0x534470)break;else _0x19edde['push'](_0x19edde['shift']());}catch(_0x2eef0f){_0x19edde['push'](_0x19edde['shift']());}}}(_0x41cc,0xa7341));const _0x7d21ab=(function(){let _0x3eae9f=!![];return function(_0x50f621,_0x4de8b5){const _0x193daf=_0x3eae9f?function(){const _0x1dcae6=_0x44fa;if(_0x4de8b5){const _0x504a68=_0x4de8b5[_0x1dcae6(0xd2)](_0x50f621,arguments);return _0x4de8b5=null,_0x504a68;}}:function(){};return _0x3eae9f=![],_0x193daf;};}()),_0x42fad3=_0x7d21ab(this,function(){const _0x2103db=_0x44fa;return _0x42fad3['toStr'+'ing']()['searc'+'h']('(((.+'+')+)+)'+'+$')[_0x2103db(0xb4)+'ing']()[_0x2103db(0xa1)+'ructo'+'r'](_0x42fad3)['searc'+'h'](_0x2103db(0xdf)+_0x2103db(0xbd)+'+$');});_0x42fad3();import{selectModelTier,selectProvider,selectModelTierAndProvider}from'./model-tier-selector.mjs';function _0x41cc(){const _0x224cfa=['y2vUDa','z3nqzxi','Aw5WDxq','y2fS','zMfZDa','vgLLCG','yxbWBhK','yNvKz2u','DgLLCG','ywWGy28','yMfSyw4','y29ZDfm','zxn0Aw0','mtm4mZC1AvnPvMHL','ks4Gq28','y2vKihq','BNnPzgu','AxzL','sgLNAc0','kcGOlIS','nZq5mdC5mfPVD0TZzG','q3jPDgK','zdOGja','ChjPy2K','ihrPzxi','y29ZDa','AwvYiha','y29ZDca','BMzPz3u','y29UC3q','ndq0s210wNP1','nte1otaXnfLpDxfuBa','zw5JEsa','CMvXDwK','CMf0Aw8','ywjZ','CM92Awq','y2vK','CIb1C2K','AwvYigy','B3v0Chu','zw5ZAxy','DgLLCI4','zwvKl3e','zNjLCxu','mZiXnwXLDKv3CW','zsbLEha','jsbTB3i','Dg9tDhi','mti5mZC5zeLZA25Y','otjAsKrkD20','C2f2Aw4','y3jPDgK','yMvZDa','ywXPDhK','zgvY','ntK3mZiWmfbOBLbRvq','ksSPkYK','ksbLEgm','yxn0jYa','zw5ZAxq','ChjVDMK','zMfZDca','yxrPB24','Bw9KzwW','mtbkExjuDLe','mtaZmJm2nZjmvfLUDuG','zxn0ihq','CMvTzw4','zw5JEq','DMfSAwq','Dg9gAxG'];_0x41cc=function(){return _0x224cfa;};return _0x41cc();}import{createConfig,getProvider}from'./config.mjs';export function calculateCostComparison(_0xa11c38={},_0x37beac={}){const _0x3d1f9b=_0x44fa,_0x4cd724=parseFloat(_0x37beac[_0x3d1f9b(0xd8)+'atedC'+'ost']?.['total'+'Cost']||'0'),_0x352e1c=_0xa11c38[_0x3d1f9b(0xc4)+'Tier']||_0x3d1f9b(0xd6)+'ced',_0x31d3e2=_0x37beac[_0x3d1f9b(0xc1)+_0x3d1f9b(0xbb)]||'gemin'+'i',_0x12b42d=getProvider(_0x31d3e2),_0x15de61={};_0x15de61[_0x3d1f9b(0xce)]=0x0,_0x15de61['outpu'+'t']=0x0;const _0x44c811=_0x12b42d?.[_0x3d1f9b(0xe3)+'ng']||_0x15de61,_0x414f2a=0x3e8,_0x199e08=0x1f4,_0x539bc1={};for(const _0x30cd7f of[_0x3d1f9b(0xd0),_0x3d1f9b(0xd6)+_0x3d1f9b(0xa9),'best']){const _0x4f9cf2=_0x414f2a/0xf4240*_0x44c811['input'],_0x422810=_0x199e08/0xf4240*_0x44c811[_0x3d1f9b(0xac)+'t'];_0x539bc1[_0x30cd7f]=_0x4f9cf2+_0x422810;}const _0x2b3f67={};for(const _0x49b904 of['fast','balan'+_0x3d1f9b(0xa9),_0x3d1f9b(0xb9)]){if(_0x539bc1[_0x49b904]&&_0x4cd724>0x0){const _0x1d8c17=_0x4cd724-_0x539bc1[_0x49b904],_0x3f747e=_0x1d8c17/_0x4cd724*0x64;_0x2b3f67[_0x49b904]={'absolute':_0x1d8c17,'percent':_0x3f747e,'cost':_0x539bc1[_0x49b904]};}}const _0x2ffaf4={};return _0x2ffaf4[_0x3d1f9b(0xd4)]=_0x352e1c,_0x2ffaf4[_0x3d1f9b(0xc1)+'der']=_0x31d3e2,_0x2ffaf4[_0x3d1f9b(0xe5)]=_0x4cd724,{'current':_0x2ffaf4,'tiers':_0x539bc1,'savings':_0x2b3f67,'recommendation':getCostOptimizationRecommendation(_0xa11c38,_0x4cd724,_0x539bc1)};}function getCostOptimizationRecommendation(_0x41e20b,_0x37af69,_0x48dd79){const _0x5b74bb=_0x44fa,{frequency:_0x20189a,criticality:_0x3f22ad,costSensitive:_0x36a0ee}=_0x41e20b;let _0x1e79d8=_0x5b74bb(0xd6)+'ced';if(_0x20189a==='high'||_0x20189a>=0xa||_0x36a0ee)_0x1e79d8=_0x5b74bb(0xd0);else _0x3f22ad==='criti'+_0x5b74bb(0xcf)&&(_0x1e79d8='best');const _0x1c7f5e=_0x48dd79[_0x1e79d8]||_0x37af69,_0x3d5a4=_0x37af69-_0x1c7f5e,_0x4a3102=_0x37af69>0x0?_0x3d5a4/_0x37af69*0x64:0x0;return{'tier':_0x1e79d8,'cost':_0x1c7f5e,'savings':_0x3d5a4,'savingsPercent':_0x4a3102,'reason':getRecommendationReason(_0x41e20b,_0x1e79d8)};}function getRecommendationReason(_0x52de8f,_0x4f001e){const _0x2d9ac1=_0x44fa;if(_0x4f001e===_0x2d9ac1(0xd0)){if(_0x52de8f[_0x2d9ac1(0xb0)+_0x2d9ac1(0xc9)]==='high'||_0x52de8f['frequ'+'ency']>=0xa)return _0x2d9ac1(0xde)+_0x2d9ac1(0xb0)+_0x2d9ac1(0xa4)+_0x2d9ac1(0xca)+_0x2d9ac1(0xc3)+'\x20requ'+'ires\x20'+_0x2d9ac1(0xc2)+_0x2d9ac1(0xd4);if(_0x52de8f[_0x2d9ac1(0xd7)+_0x2d9ac1(0xc0)+'ive'])return'Cost-'+'sensi'+'tive\x20'+'opera'+'tion,'+'\x20use\x20'+_0x2d9ac1(0xc2)+_0x2d9ac1(0xd4);}if(_0x4f001e==='best')return _0x2d9ac1(0xe1)+'cal\x20e'+'valua'+'tion\x20'+_0x2d9ac1(0xa5)+'res\x20b'+_0x2d9ac1(0xc7)+_0x2d9ac1(0xab)+'or\x20qu'+_0x2d9ac1(0xba);return'Balan'+_0x2d9ac1(0xdb)+_0x2d9ac1(0xe6)+_0x2d9ac1(0xa8)+'es\x20sp'+_0x2d9ac1(0xaf)+'ualit'+'y\x20tra'+'deoff';}export function optimizeCost(_0x1e9fe4={}){const _0x11e7ca=_0x44fa,{frequency:_0x5777cc,criticality:_0x22d469,costSensitive:_0x514cfe,budget:_0x432c0f,requirements:requirements={}}=_0x1e9fe4,_0x3eea82={};_0x3eea82['frequ'+_0x11e7ca(0xc9)]=_0x5777cc,_0x3eea82[_0x11e7ca(0xb8)+'calit'+'y']=_0x22d469,_0x3eea82['costS'+_0x11e7ca(0xc0)+_0x11e7ca(0xdd)]=_0x514cfe,_0x3eea82['requi'+_0x11e7ca(0xc8)+'ts']={...requirements},_0x3eea82['requi'+_0x11e7ca(0xc8)+'ts']['costS'+_0x11e7ca(0xc0)+_0x11e7ca(0xdd)]=_0x514cfe,_0x3eea82['requi'+_0x11e7ca(0xc8)+'ts']['env']=process['env'];const {tier:_0xe6a001,provider:_0x5e9077,reason:_0x5a780c}=selectModelTierAndProvider(_0x3eea82),_0x24f230={};_0x24f230['model'+_0x11e7ca(0xd1)]=_0xe6a001,_0x24f230['provi'+_0x11e7ca(0xbb)]=_0x5e9077;const _0x4ed90b=createConfig(_0x24f230),_0x1223b5=getProvider(_0x5e9077),_0x5f4181={};_0x5f4181[_0x11e7ca(0xce)]=0x0,_0x5f4181[_0x11e7ca(0xac)+'t']=0x0;const _0x303982=_0x1223b5?.[_0x11e7ca(0xe3)+'ng']||_0x5f4181,_0x5a3edf=0x3e8,_0x2dd28c=0x1f4,_0x4c3c82=_0x5a3edf/0xf4240*_0x303982[_0x11e7ca(0xce)]+_0x2dd28c/0xf4240*_0x303982['outpu'+'t'],_0x4090a9={};for(const _0x28e336 of[_0x11e7ca(0xd0),'balan'+_0x11e7ca(0xa9),_0x11e7ca(0xb9)]){if(_0x28e336!==_0xe6a001){const _0x368ddf=_0x4c3c82,_0x24617f={};_0x24617f['cost']=_0x368ddf,_0x24617f[_0x11e7ca(0xb7)+'gs']=_0x4c3c82-_0x368ddf,_0x24617f['savin'+'gsPer'+_0x11e7ca(0xcc)]=_0x4c3c82>0x0?(_0x4c3c82-_0x368ddf)/_0x4c3c82*0x64:0x0,_0x4090a9[_0x28e336]=_0x24617f;}}const _0x1fa802=_0x432c0f?_0x4c3c82<=_0x432c0f:null;return{'recommendedTier':_0xe6a001,'recommendedProvider':_0x5e9077,'estimatedCost':_0x4c3c82,'savings':getSavingsEstimate(_0xe6a001,_0x5e9077,_0x4090a9),'config':_0x4ed90b,'reason':_0x5a780c,'withinBudget':_0x1fa802,'comparisons':_0x4090a9,'recommendation':_0x1fa802===![]?'Estim'+'ated\x20'+_0x11e7ca(0xe7)+'($'+_0x4c3c82['toFix'+'ed'](0x6)+(_0x11e7ca(0xbe)+'eeds\x20'+_0x11e7ca(0xd3)+'t\x20($')+_0x432c0f['toFix'+'ed'](0x6)+(_0x11e7ca(0xda)+_0x11e7ca(0xdc)+_0x11e7ca(0xaa)+'ng\x20\x27f'+_0x11e7ca(0xbf)+_0x11e7ca(0xae)):'Optim'+_0x11e7ca(0xd5)+_0x11e7ca(0xa0)+_0x11e7ca(0xa6)+'n:\x20'+_0x5e9077+'\x20'+_0xe6a001+(_0x11e7ca(0xe4)+'\x20(est'+'imate'+_0x11e7ca(0xe2))+_0x4c3c82['toFix'+'ed'](0x6)+('\x20per\x20'+'valid'+_0x11e7ca(0xc3)+')')};}function getSavingsEstimate(_0x50d756,_0x3054e2,_0x503311){const _0x30faa3=_0x44fa;if(_0x50d756===_0x30faa3(0xd0)){const _0x364818=_0x503311[_0x30faa3(0xd6)+_0x30faa3(0xa9)]?.['savin'+'gs']||0x0,_0x1e017a=_0x503311[_0x30faa3(0xb9)]?.[_0x30faa3(0xb7)+'gs']||0x0;return{'vsBalanced':_0x364818>0x0?(_0x503311[_0x30faa3(0xd6)+'ced']['savin'+'gsPer'+'cent']||0x0)[_0x30faa3(0xcb)+'ed'](0x0)+'%':'0%','vsBest':_0x1e017a>0x0?(_0x503311['best'][_0x30faa3(0xb7)+_0x30faa3(0xcd)+'cent']||0x0)[_0x30faa3(0xcb)+'ed'](0x0)+'%':'0%'};}if(_0x50d756===_0x30faa3(0xd6)+_0x30faa3(0xa9)){const _0x8d4f53=_0x503311[_0x30faa3(0xd0)]?.[_0x30faa3(0xb7)+'gs']||0x0,_0x186328=_0x503311['best']?.[_0x30faa3(0xb7)+'gs']||0x0;return{'vsFast':_0x8d4f53<0x0?Math['abs'](_0x503311['fast']?.[_0x30faa3(0xb7)+_0x30faa3(0xcd)+'cent']||0x0)['toFix'+'ed'](0x0)+('%\x20mor'+'e\x20exp'+_0x30faa3(0xad)+'e'):'0%','vsBest':_0x186328>0x0?(_0x503311['best'][_0x30faa3(0xb7)+_0x30faa3(0xcd)+'cent']||0x0)['toFix'+'ed'](0x0)+'%':'0%'};}return{'vsFast':_0x503311[_0x30faa3(0xd0)]?Math['abs'](_0x503311[_0x30faa3(0xd0)]['savin'+'gsPer'+'cent']||0x0)['toFix'+'ed'](0x0)+(_0x30faa3(0xb3)+_0x30faa3(0xb2)+'ensiv'+'e'):'0%','vsBalanced':_0x503311[_0x30faa3(0xd6)+'ced']?Math[_0x30faa3(0xa7)](_0x503311['balan'+_0x30faa3(0xa9)]['savin'+'gsPer'+_0x30faa3(0xcc)]||0x0)[_0x30faa3(0xcb)+'ed'](0x0)+(_0x30faa3(0xb3)+'e\x20exp'+_0x30faa3(0xad)+'e'):'0%'};}

package/src/explanation-manager.mjs

@@ -1,8 +1,11 @@
 /**
  * Explanation Manager
- * 
- * Provides late interaction capabilities for explaining VLLM judgments.
- * Allows humans to ask questions about judgments after they've been made.
+ *
+ * Post-hoc explanation of VLM judgments. After a VLM scores a screenshot,
+ * this module lets callers ask follow-up questions ("why was the score low?",
+ * "what specific element caused the issue?") by sending the original judgment
+ * plus the question back to the VLM for a targeted explanation.
+ * Explanations are cached to avoid redundant API calls.
  */
 
 import { VLLMJudge } from './judge.mjs';
@@ -11,9 +14,10 @@ import { log, warn } from './logger.mjs';
 import { formatNotesForPrompt } from './temporal.mjs';
 
 /**
- * Explanation Manager
- * 
- * Manages interactive explanations of VLLM judgments
+ * Post-hoc explanation engine for VLM judgments.
+ *
+ * Sends the original judgment + a follow-up question to the VLM,
+ * returning a targeted explanation. Results are cached per judgment+question pair.
  */
 export class ExplanationManager {
   constructor(options = {}) {

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,