Wiki source code of POC Requirements (POC1 & POC2)
Last modified by Robert Schaub on 2025/12/24 21:27
Show last authors
| author | version | line-number | content |
|---|---|---|---|
| 1 | = POC Requirements = | ||
| 2 | |||
| 3 | |||
| 4 | {{info}} | ||
| 5 | **POC1 Architecture:** 3-stage AKEL pipeline (Extract → Analyze → Holistic) with Redis caching, credit tracking, and LLM abstraction layer. | ||
| 6 | |||
| 7 | See [[POC1 API Specification>>FactHarbor.Specification.POC.API-and-Schemas.WebHome]] for complete technical details. | ||
| 8 | {{/info}} | ||
| 9 | |||
| 10 | |||
| 11 | |||
| 12 | **Status:** ✅ Approved for Development | ||
| 13 | **Version:** 2.0 (Updated after Specification Cross-Check) | ||
| 14 | **Goal:** Prove that AI can extract claims and determine verdicts automatically without human intervention | ||
| 15 | |||
| 16 | == 1. POC Overview == | ||
| 17 | |||
| 18 | === 1.1 What POC Tests === | ||
| 19 | |||
| 20 | **Core Question:** | ||
| 21 | > Can AI automatically extract factual claims from articles and evaluate them with reasonable verdicts? | ||
| 22 | |||
| 23 | **What we're proving:** | ||
| 24 | * AI can identify factual claims from text | ||
| 25 | * AI can evaluate those claims and produce verdicts | ||
| 26 | * Output is comprehensible and useful | ||
| 27 | * Fully automated approach is viable | ||
| 28 | |||
| 29 | **What we're NOT testing:** | ||
| 30 | * Scenario generation (deferred to POC2) | ||
| 31 | * Evidence display (deferred to POC2) | ||
| 32 | * Production scalability | ||
| 33 | * Perfect accuracy | ||
| 34 | * Complete feature set | ||
| 35 | |||
| 36 | === 1.2 Scenarios Deferred to POC2 === | ||
| 37 | |||
| 38 | **Intentional Simplification:** | ||
| 39 | |||
| 40 | Scenarios are a core component of the full FactHarbor system (Claims → Scenarios → Evidence → Verdicts), but are **deliberately excluded from POC1**. | ||
| 41 | |||
| 42 | **Rationale:** | ||
| 43 | * **POC1 tests:** Can AI extract claims and generate verdicts? | ||
| 44 | * **POC2 will add:** Scenario generation and management | ||
| 45 | * **Open questions remain:** Should scenarios be separate entities? How are they sequenced with evidence gathering? What's the optimal workflow? | ||
| 46 | |||
| 47 | **Design Decision:** | ||
| 48 | |||
| 49 | Prove basic AI capability first, then add scenario complexity based on POC1 learnings. This is good engineering: test the hardest part (AI fact-checking) before adding architectural complexity. | ||
| 50 | |||
| 51 | **No Risk:** | ||
| 52 | |||
| 53 | Scenarios are additive complexity, not foundational. Deferring them to POC2 allows: | ||
| 54 | * Faster POC1 validation | ||
| 55 | * Learning from POC1 to inform scenario design | ||
| 56 | * Iterative approach: fail fast if basic AI doesn't work | ||
| 57 | * Flexibility to adjust scenario architecture based on POC1 insights | ||
| 58 | |||
| 59 | **Full System Workflow (Future):** | ||
| 60 | {{code}} | ||
| 61 | Claims → Scenarios → Evidence → Verdicts | ||
| 62 | {{/code}} | ||
| 63 | |||
| 64 | **POC1 Simplified Workflow:** | ||
| 65 | {{code}} | ||
| 66 | Claims → Verdicts (scenarios implicit in reasoning) | ||
| 67 | {{/code}} | ||
| 68 | |||
| 69 | == 2. POC Output Specification == | ||
| 70 | |||
| 71 | === 2.1 Component 1: ANALYSIS SUMMARY (Context-Aware) === | ||
| 72 | |||
| 73 | **What:** Context-aware overview that considers both individual claims AND their relationship to the article's main argument | ||
| 74 | |||
| 75 | **Length:** 4-6 sentences | ||
| 76 | |||
| 77 | **Content (Required Elements):** | ||
| 78 | 1. **Article's main thesis/claim** - What is the article trying to argue or prove? | ||
| 79 | 2. **Claim count and verdicts** - How many claims analyzed, distribution of verdicts | ||
| 80 | 3. **Central vs. supporting claims** - Which claims are central to the article's argument? | ||
| 81 | 4. **Relationship assessment** - Do the claims support the article's conclusion? | ||
| 82 | 5. **Overall credibility** - Final assessment considering claim importance | ||
| 83 | |||
| 84 | **Critical Innovation:** | ||
| 85 | |||
| 86 | POC1 tests whether AI can understand that **article credibility ≠ simple average of claim verdicts**. An article might: | ||
| 87 | * Make accurate supporting facts but draw unsupported conclusions | ||
| 88 | * Have one false central claim that invalidates the whole argument | ||
| 89 | * Misframe accurate information to mislead | ||
| 90 | |||
| 91 | **Good Example (Context-Aware):** | ||
| 92 | {{code}} | ||
| 93 | This article argues that coffee cures cancer based on its antioxidant | ||
| 94 | content. We analyzed 3 factual claims: 2 about coffee's chemical | ||
| 95 | properties are well-supported, but the main causal claim is refuted | ||
| 96 | by current evidence. The article confuses correlation with causation. | ||
| 97 | Overall assessment: MISLEADING - makes an unsupported medical claim | ||
| 98 | despite citing some accurate facts. | ||
| 99 | {{/code}} | ||
| 100 | |||
| 101 | **Poor Example (Simple Aggregation - Don't Do This):** | ||
| 102 | {{code}} | ||
| 103 | This article makes 3 claims. 2 are well-supported and 1 is refuted. | ||
| 104 | Overall assessment: mostly accurate (67% accurate). | ||
| 105 | {{/code}} | ||
| 106 | ↑ This misses that the refuted claim IS the article's main point! | ||
| 107 | |||
| 108 | **What POC1 Tests:** | ||
| 109 | |||
| 110 | Can AI identify and assess: | ||
| 111 | * ✅ The article's main thesis/conclusion? | ||
| 112 | * ✅ Which claims are central vs. supporting? | ||
| 113 | * ✅ Whether the evidence supports the conclusion? | ||
| 114 | * ✅ Overall credibility considering logical structure? | ||
| 115 | |||
| 116 | **If AI Cannot Do This:** | ||
| 117 | |||
| 118 | That's valuable to learn in POC1! We'll: | ||
| 119 | * Note as limitation | ||
| 120 | * Fall back to simple aggregation with warning | ||
| 121 | * Design explicit article-level analysis for POC2 | ||
| 122 | |||
| 123 | === 2.2 Component 2: CLAIMS IDENTIFICATION === | ||
| 124 | |||
| 125 | **What:** List of factual claims extracted from article | ||
| 126 | **Format:** Numbered list | ||
| 127 | **Quantity:** 3-5 claims | ||
| 128 | **Requirements:** | ||
| 129 | * Factual claims only (not opinions/questions) | ||
| 130 | * Clearly stated | ||
| 131 | * Automatically extracted by AI | ||
| 132 | |||
| 133 | **Example:** | ||
| 134 | {{code}} | ||
| 135 | CLAIMS IDENTIFIED: | ||
| 136 | |||
| 137 | [1] Coffee reduces diabetes risk by 30% | ||
| 138 | [2] Coffee improves heart health | ||
| 139 | [3] Decaf has same benefits as regular | ||
| 140 | [4] Coffee prevents Alzheimer's completely | ||
| 141 | {{/code}} | ||
| 142 | |||
| 143 | === 2.3 Component 3: CLAIMS VERDICTS === | ||
| 144 | |||
| 145 | **What:** Verdict for each claim identified | ||
| 146 | **Format:** Per claim structure | ||
| 147 | |||
| 148 | **Required Elements:** | ||
| 149 | * **Verdict Label:** WELL-SUPPORTED / PARTIALLY SUPPORTED / UNCERTAIN / REFUTED | ||
| 150 | * **Confidence Score:** 0-100% | ||
| 151 | * **Brief Reasoning:** 1-3 sentences explaining why | ||
| 152 | * **Risk Tier:** A (High) / B (Medium) / C (Low) - for demonstration | ||
| 153 | |||
| 154 | **Example:** | ||
| 155 | {{code}} | ||
| 156 | VERDICTS: | ||
| 157 | |||
| 158 | [1] WELL-SUPPORTED (85%) [Risk: C] | ||
| 159 | Multiple studies confirm 25-30% risk reduction with regular consumption. | ||
| 160 | |||
| 161 | [2] UNCERTAIN (65%) [Risk: B] | ||
| 162 | Evidence is mixed. Some studies show benefits, others show no effect. | ||
| 163 | |||
| 164 | [3] PARTIALLY SUPPORTED (60%) [Risk: C] | ||
| 165 | Some benefits overlap, but caffeine-related benefits are reduced in decaf. | ||
| 166 | |||
| 167 | [4] REFUTED (90%) [Risk: B] | ||
| 168 | No evidence for complete prevention. Claim is significantly overstated. | ||
| 169 | {{/code}} | ||
| 170 | |||
| 171 | **Risk Tier Display:** | ||
| 172 | * **Tier A (Red):** High Risk - Medical/Legal/Safety/Elections | ||
| 173 | * **Tier B (Yellow):** Medium Risk - Policy/Science/Causality | ||
| 174 | * **Tier C (Green):** Low Risk - Facts/Definitions/History | ||
| 175 | |||
| 176 | **Note:** Risk tier shown for demonstration purposes in POC. Full system uses risk tiers to determine review workflow. | ||
| 177 | |||
| 178 | === 2.4 Component 4: ARTICLE SUMMARY (Optional) === | ||
| 179 | |||
| 180 | **What:** Brief summary of original article content | ||
| 181 | **Length:** 3-5 sentences | ||
| 182 | **Tone:** Neutral (article's position, not FactHarbor's analysis) | ||
| 183 | |||
| 184 | **Example:** | ||
| 185 | {{code}} | ||
| 186 | ARTICLE SUMMARY: | ||
| 187 | |||
| 188 | Health News Today article discusses coffee benefits, citing studies | ||
| 189 | on diabetes and Alzheimer's. Author highlights research linking coffee | ||
| 190 | to disease prevention. Recommends 2-3 cups daily for optimal health. | ||
| 191 | {{/code}} | ||
| 192 | |||
| 193 | === 2.5 Component 5: USAGE STATISTICS (Cost Tracking) === | ||
| 194 | |||
| 195 | **What:** LLM usage metrics for cost optimization and scaling decisions | ||
| 196 | |||
| 197 | **Purpose:** | ||
| 198 | * Understand cost per analysis | ||
| 199 | * Identify optimization opportunities | ||
| 200 | * Project costs at scale | ||
| 201 | * Inform architecture decisions | ||
| 202 | |||
| 203 | **Display Format:** | ||
| 204 | {{code}} | ||
| 205 | USAGE STATISTICS: | ||
| 206 | • Article: 2,450 words (12,300 characters) | ||
| 207 | • Input tokens: 15,234 | ||
| 208 | • Output tokens: 892 | ||
| 209 | • Total tokens: 16,126 | ||
| 210 | • Estimated cost: $0.24 USD | ||
| 211 | • Response time: 8.3 seconds | ||
| 212 | • Cost per claim: $0.048 | ||
| 213 | • Model: claude-sonnet-4-20250514 | ||
| 214 | {{/code}} | ||
| 215 | |||
| 216 | **Why This Matters:** | ||
| 217 | |||
| 218 | At scale, LLM costs are critical: | ||
| 219 | * 10,000 articles/month ≈ $200-500/month | ||
| 220 | * 100,000 articles/month ≈ $2,000-5,000/month | ||
| 221 | * Cost optimization can reduce expenses 30-50% | ||
| 222 | |||
| 223 | **What POC1 Learns:** | ||
| 224 | * How cost scales with article length | ||
| 225 | * Prompt optimization opportunities (caching, compression) | ||
| 226 | * Output verbosity tradeoffs | ||
| 227 | * Model selection strategy (FAST vs. REASONING roles) | ||
| 228 | * Article length limits (if needed) | ||
| 229 | |||
| 230 | **Implementation:** | ||
| 231 | * Claude API already returns usage data | ||
| 232 | * No extra API calls needed | ||
| 233 | * Display to user + log for aggregate analysis | ||
| 234 | * Test with articles of varying lengths | ||
| 235 | |||
| 236 | **Critical for GO/NO-GO:** Unit economics must be viable at scale! | ||
| 237 | |||
| 238 | === 2.6 Total Output Size === | ||
| 239 | |||
| 240 | **Combined:** ~220-350 words | ||
| 241 | * Analysis Summary (Context-Aware): 60-90 words (4-6 sentences) | ||
| 242 | * Claims Identification: 30-50 words | ||
| 243 | * Claims Verdicts: 100-150 words | ||
| 244 | * Article Summary: 30-50 words (optional) | ||
| 245 | |||
| 246 | **Note:** Analysis summary is slightly longer (4-6 sentences vs. 3-5) to accommodate context-aware assessment of article structure and logical reasoning. | ||
| 247 | |||
| 248 | == 3. What's NOT in POC Scope == | ||
| 249 | |||
| 250 | === 3.1 Feature Exclusions === | ||
| 251 | |||
| 252 | The following are **explicitly excluded** from POC: | ||
| 253 | |||
| 254 | **Content Features:** | ||
| 255 | * ❌ Scenarios (deferred to POC2) | ||
| 256 | * ❌ Evidence display (supporting/opposing lists) | ||
| 257 | * ❌ Source links (clickable references) | ||
| 258 | * ❌ Detailed reasoning chains | ||
| 259 | * ❌ Source quality ratings (shown but not detailed) | ||
| 260 | * ❌ Contradiction detection (basic only) | ||
| 261 | * ❌ Risk assessment (shown but not workflow-integrated) | ||
| 262 | |||
| 263 | **Platform Features:** | ||
| 264 | * ❌ User accounts / authentication | ||
| 265 | * ❌ Saved history | ||
| 266 | * ❌ Search functionality | ||
| 267 | * ❌ Claim comparison | ||
| 268 | * ❌ User contributions | ||
| 269 | * ❌ Commenting system | ||
| 270 | * ❌ Social sharing | ||
| 271 | |||
| 272 | **Technical Features:** | ||
| 273 | * ❌ Browser extensions | ||
| 274 | * ❌ Mobile apps | ||
| 275 | * ❌ API endpoints | ||
| 276 | * ❌ Webhooks | ||
| 277 | * ❌ Export features (PDF, CSV) | ||
| 278 | |||
| 279 | **Quality Features:** | ||
| 280 | * ❌ Accessibility (WCAG compliance) | ||
| 281 | * ❌ Multilingual support | ||
| 282 | * ❌ Mobile optimization | ||
| 283 | * ❌ Media verification (images/videos) | ||
| 284 | |||
| 285 | **Production Features:** | ||
| 286 | * ❌ Security hardening | ||
| 287 | * ❌ Privacy compliance (GDPR) | ||
| 288 | * ❌ Terms of service | ||
| 289 | * ❌ Monitoring/logging | ||
| 290 | * ❌ Error tracking | ||
| 291 | * ❌ Analytics | ||
| 292 | * ❌ A/B testing | ||
| 293 | |||
| 294 | == 4. POC Simplifications vs. Full System == | ||
| 295 | |||
| 296 | === 4.1 Architecture Comparison === | ||
| 297 | |||
| 298 | **POC Architecture (Simplified):** | ||
| 299 | {{code}} | ||
| 300 | User Input → Single AKEL Call → Output Display | ||
| 301 | (all processing) | ||
| 302 | {{/code}} | ||
| 303 | |||
| 304 | **Full System Architecture:** | ||
| 305 | {{code}} | ||
| 306 | User Input → Claim Extractor → Claim Classifier → Scenario Generator | ||
| 307 | → Evidence Summarizer → Contradiction Detector → Verdict Generator | ||
| 308 | → Quality Gates → Publication → Output Display | ||
| 309 | {{/code}} | ||
| 310 | |||
| 311 | **Key Differences:** | ||
| 312 | |||
| 313 | |=Aspect|=POC1|=Full System | ||
| 314 | |Processing|Single API call|Multi-component pipeline | ||
| 315 | |Scenarios|None (implicit)|Explicit entities with versioning | ||
| 316 | |Evidence|Basic retrieval|Comprehensive with quality scoring | ||
| 317 | |Quality Gates|Simplified (4 basic checks)|Full validation infrastructure | ||
| 318 | |Workflow|3 steps (input/process/output)|6 phases with gates | ||
| 319 | |Data Model|Stateless (no database)|PostgreSQL + Redis + S3 | ||
| 320 | |Architecture|Single prompt to Claude|AKEL Orchestrator + Components | ||
| 321 | |||
| 322 | === 4.2 Workflow Comparison === | ||
| 323 | |||
| 324 | **POC1 Workflow:** | ||
| 325 | 1. User submits text/URL | ||
| 326 | 2. Single AKEL call (all processing in one prompt) | ||
| 327 | 3. Display results | ||
| 328 | **Total: 3 steps, ~10-18 seconds** | ||
| 329 | |||
| 330 | **Full System Workflow:** | ||
| 331 | 1. **Claim Submission** (extraction, normalization, clustering) | ||
| 332 | 2. **Scenario Building** (definitions, assumptions, boundaries) | ||
| 333 | 3. **Evidence Handling** (retrieval, assessment, linking) | ||
| 334 | 4. **Verdict Creation** (synthesis, reasoning, approval) | ||
| 335 | 5. **Public Presentation** (summaries, landscapes, deep dives) | ||
| 336 | 6. **Time Evolution** (versioning, re-evaluation triggers) | ||
| 337 | **Total: 6 phases with quality gates, ~10-30 seconds** | ||
| 338 | |||
| 339 | === 4.3 Why POC is Simplified === | ||
| 340 | |||
| 341 | **Engineering Rationale:** | ||
| 342 | |||
| 343 | 1. **Test core capability first:** Can AI do basic fact-checking without humans? | ||
| 344 | 2. **Fail fast:** If AI can't generate reasonable verdicts, pivot early | ||
| 345 | 3. **Learn before building:** POC1 insights inform full architecture | ||
| 346 | 4. **Iterative approach:** Add complexity only after validating foundations | ||
| 347 | 5. **Resource efficiency:** Don't build full system if core concept fails | ||
| 348 | |||
| 349 | **Acceptable Trade-offs:** | ||
| 350 | |||
| 351 | * ✅ POC proves AI capability (most risky assumption) | ||
| 352 | * ✅ POC validates user comprehension (can people understand output?) | ||
| 353 | * ❌ POC doesn't validate full workflow (test in Beta) | ||
| 354 | * ❌ POC doesn't validate scale (test in Beta) | ||
| 355 | * ❌ POC doesn't validate scenario architecture (design in POC2) | ||
| 356 | |||
| 357 | === 4.4 Gap Between POC1 and POC2/Beta === | ||
| 358 | |||
| 359 | **What needs to be built for POC2:** | ||
| 360 | * Scenario generation component | ||
| 361 | * Evidence Model structure (full) | ||
| 362 | * Scenario-evidence linking | ||
| 363 | * Multi-interpretation comparison | ||
| 364 | * Truth landscape visualization | ||
| 365 | |||
| 366 | **What needs to be built for Beta:** | ||
| 367 | * Multi-component AKEL pipeline | ||
| 368 | * Quality gate infrastructure | ||
| 369 | * Review workflow system | ||
| 370 | * Audit sampling framework | ||
| 371 | * Production data model | ||
| 372 | * Federation architecture (Release 1.0) | ||
| 373 | |||
| 374 | **POC1 → POC2 is significant architectural expansion.** | ||
| 375 | |||
| 376 | == 5. Publication Mode & Labeling == | ||
| 377 | |||
| 378 | === 5.1 POC Publication Mode === | ||
| 379 | |||
| 380 | **Mode:** Mode 2 (AI-Generated, No Prior Human Review) | ||
| 381 | |||
| 382 | Per FactHarbor Specification Section 11 "POC v1 Behavior": | ||
| 383 | * Produces public AI-generated output | ||
| 384 | * No human approval gate | ||
| 385 | * Clear AI-Generated labeling | ||
| 386 | * All quality gates active (simplified) | ||
| 387 | * Risk tier classification shown (demo) | ||
| 388 | |||
| 389 | === 5.2 User-Facing Labels === | ||
| 390 | |||
| 391 | **Primary Label (top of analysis):** | ||
| 392 | {{code}} | ||
| 393 | ╔════════════════════════════════════════════════════════════╗ | ||
| 394 | ║ [AI-GENERATED - POC/DEMO] ║ | ||
| 395 | ║ ║ | ||
| 396 | ║ This analysis was produced entirely by AI and has not ║ | ||
| 397 | ║ been human-reviewed. Use for demonstration purposes. ║ | ||
| 398 | ║ ║ | ||
| 399 | ║ Source: AI/AKEL v1.0 (POC) ║ | ||
| 400 | ║ Review Status: Not Reviewed (Proof-of-Concept) ║ | ||
| 401 | ║ Quality Gates: 4/4 Passed (Simplified) ║ | ||
| 402 | ║ Last Updated: [timestamp] ║ | ||
| 403 | ╚════════════════════════════════════════════════════════════╝ | ||
| 404 | {{/code}} | ||
| 405 | |||
| 406 | **Per-Claim Risk Labels:** | ||
| 407 | * **[Risk: A]** 🔴 High Risk (Medical/Legal/Safety) | ||
| 408 | * **[Risk: B]** 🟡 Medium Risk (Policy/Science) | ||
| 409 | * **[Risk: C]** 🟢 Low Risk (Facts/Definitions) | ||
| 410 | |||
| 411 | === 5.3 Display Requirements === | ||
| 412 | |||
| 413 | **Must Show:** | ||
| 414 | * AI-Generated status (prominent) | ||
| 415 | * POC/Demo disclaimer | ||
| 416 | * Risk tier per claim | ||
| 417 | * Confidence scores (0-100%) | ||
| 418 | * Quality gate status (passed/failed) | ||
| 419 | * Timestamp | ||
| 420 | |||
| 421 | **Must NOT Claim:** | ||
| 422 | * Human review | ||
| 423 | * Production quality | ||
| 424 | * Medical/legal advice | ||
| 425 | * Authoritative verdicts | ||
| 426 | * Complete accuracy | ||
| 427 | |||
| 428 | === 5.4 Mode 2 vs. Full System Publication === | ||
| 429 | |||
| 430 | |=Element|=POC Mode 2|=Full System Mode 2|=Full System Mode 3 | ||
| 431 | |Label|AI-Generated (POC)|AI-Generated|AKEL-Generated | ||
| 432 | |Review|None|None|Human-Reviewed | ||
| 433 | |Quality Gates|4 (simplified)|6 (full)|6 (full) + Human | ||
| 434 | |Audit|None (POC)|Sampling (5-50%)|Pre-publication | ||
| 435 | |Risk Display|Demo only|Workflow-integrated|Validated | ||
| 436 | |User Actions|View only|Flag for review|Trust rating | ||
| 437 | |||
| 438 | == 6. Quality Gates (Simplified Implementation) == | ||
| 439 | |||
| 440 | === 6.1 Overview === | ||
| 441 | |||
| 442 | Per FactHarbor Specification Section 6, all AI-generated content must pass quality gates before publication. POC implements **simplified versions** of the 4 mandatory gates. | ||
| 443 | |||
| 444 | **Full System Has 4 Gates:** | ||
| 445 | 1. Source Quality | ||
| 446 | 2. Contradiction Search (MANDATORY) | ||
| 447 | 3. Uncertainty Quantification | ||
| 448 | 4. Structural Integrity | ||
| 449 | |||
| 450 | **POC Implements Simplified Versions:** | ||
| 451 | * Focus on demonstrating concept | ||
| 452 | * Basic implementations sufficient | ||
| 453 | * Failures displayed to user (not blocking) | ||
| 454 | * Full system has comprehensive validation | ||
| 455 | |||
| 456 | === 6.2 Gate 1: Source Quality (Basic) === | ||
| 457 | |||
| 458 | **Full System Requirements:** | ||
| 459 | * Primary sources identified and accessible | ||
| 460 | * Source reliability scored against whitelist | ||
| 461 | * Citation completeness verified | ||
| 462 | * Publication dates checked | ||
| 463 | * Author credentials validated | ||
| 464 | |||
| 465 | **POC Implementation:** | ||
| 466 | * ✅ At least 2 sources found | ||
| 467 | * ✅ Sources accessible (URLs valid) | ||
| 468 | * ❌ No whitelist checking | ||
| 469 | * ❌ No credential validation | ||
| 470 | * ❌ No comprehensive reliability scoring | ||
| 471 | |||
| 472 | **Pass Criteria:** ≥2 accessible sources found | ||
| 473 | |||
| 474 | **Failure Handling:** Display error message, don't generate verdict | ||
| 475 | |||
| 476 | === 6.3 Gate 2: Contradiction Search (Basic) === | ||
| 477 | |||
| 478 | **Full System Requirements:** | ||
| 479 | * Counter-evidence actively searched | ||
| 480 | * Reservations and limitations identified | ||
| 481 | * Alternative interpretations explored | ||
| 482 | * Bubble detection (echo chambers, conspiracy theories) | ||
| 483 | * Cross-cultural and international perspectives | ||
| 484 | * Academic literature (supporting AND opposing) | ||
| 485 | |||
| 486 | **POC Implementation:** | ||
| 487 | * ✅ Basic search for counter-evidence | ||
| 488 | * ✅ Identify obvious contradictions | ||
| 489 | * ❌ No comprehensive academic search | ||
| 490 | * ❌ No bubble detection | ||
| 491 | * ❌ No systematic alternative interpretation search | ||
| 492 | * ❌ No international perspective verification | ||
| 493 | |||
| 494 | **Pass Criteria:** Basic contradiction search attempted | ||
| 495 | |||
| 496 | **Failure Handling:** Note "limited contradiction search" in output | ||
| 497 | |||
| 498 | === 6.4 Gate 3: Uncertainty Quantification (Basic) === | ||
| 499 | |||
| 500 | **Full System Requirements:** | ||
| 501 | * Confidence scores calculated for all claims/verdicts | ||
| 502 | * Limitations explicitly stated | ||
| 503 | * Data gaps identified and disclosed | ||
| 504 | * Strength of evidence assessed | ||
| 505 | * Alternative scenarios considered | ||
| 506 | |||
| 507 | **POC Implementation:** | ||
| 508 | * ✅ Confidence scores (0-100%) | ||
| 509 | * ✅ Basic uncertainty acknowledgment | ||
| 510 | * ❌ No detailed limitation disclosure | ||
| 511 | * ❌ No data gap identification | ||
| 512 | * ❌ No alternative scenario consideration (deferred to POC2) | ||
| 513 | |||
| 514 | **Pass Criteria:** Confidence score assigned | ||
| 515 | |||
| 516 | **Failure Handling:** Show "Confidence: Unknown" if calculation fails | ||
| 517 | |||
| 518 | === 6.5 Gate 4: Structural Integrity (Basic) === | ||
| 519 | |||
| 520 | **Full System Requirements:** | ||
| 521 | * No hallucinations detected (fact-checking against sources) | ||
| 522 | * Logic chain valid and traceable | ||
| 523 | * References accessible and verifiable | ||
| 524 | * No circular reasoning | ||
| 525 | * Premises clearly stated | ||
| 526 | |||
| 527 | **POC Implementation:** | ||
| 528 | * ✅ Basic coherence check | ||
| 529 | * ✅ References accessible | ||
| 530 | * ❌ No comprehensive hallucination detection | ||
| 531 | * ❌ No formal logic validation | ||
| 532 | * ❌ No premise extraction and verification | ||
| 533 | |||
| 534 | **Pass Criteria:** Output is coherent and references are accessible | ||
| 535 | |||
| 536 | **Failure Handling:** Display error message | ||
| 537 | |||
| 538 | === 6.6 Quality Gate Display === | ||
| 539 | |||
| 540 | **POC shows simplified status:** | ||
| 541 | {{code}} | ||
| 542 | Quality Gates: 4/4 Passed (Simplified) | ||
| 543 | ✓ Source Quality: 3 sources found | ||
| 544 | ✓ Contradiction Search: Basic search completed | ||
| 545 | ✓ Uncertainty: Confidence scores assigned | ||
| 546 | ✓ Structural Integrity: Output coherent | ||
| 547 | {{/code}} | ||
| 548 | |||
| 549 | **If any gate fails:** | ||
| 550 | {{code}} | ||
| 551 | Quality Gates: 3/4 Passed (Simplified) | ||
| 552 | ✓ Source Quality: 3 sources found | ||
| 553 | ✗ Contradiction Search: Search failed - limited evidence | ||
| 554 | ✓ Uncertainty: Confidence scores assigned | ||
| 555 | ✓ Structural Integrity: Output coherent | ||
| 556 | |||
| 557 | Note: This analysis has limited evidence. Use with caution. | ||
| 558 | {{/code}} | ||
| 559 | |||
| 560 | === 6.7 Simplified vs. Full System === | ||
| 561 | |||
| 562 | |=Gate|=POC (Simplified)|=Full System | ||
| 563 | |Source Quality|≥2 sources accessible|Whitelist scoring, credentials, comprehensiveness | ||
| 564 | |Contradiction|Basic search|Systematic academic + media + international | ||
| 565 | |Uncertainty|Confidence % assigned|Detailed limitations, data gaps, alternatives | ||
| 566 | |Structural|Coherence check|Hallucination detection, logic validation, premise check | ||
| 567 | |||
| 568 | **POC Goal:** Demonstrate that quality gates are possible, not perfect implementation. | ||
| 569 | |||
| 570 | == 7. AKEL Architecture Comparison == | ||
| 571 | |||
| 572 | === 7.1 POC AKEL (Simplified) === | ||
| 573 | |||
| 574 | **Implementation:** | ||
| 575 | * Single provider API call (REASONING model) | ||
| 576 | * One comprehensive prompt | ||
| 577 | * All processing in single request | ||
| 578 | * No separate components | ||
| 579 | * No orchestration layer | ||
| 580 | |||
| 581 | **Prompt Structure:** | ||
| 582 | {{code}} | ||
| 583 | Task: Analyze this article and provide: | ||
| 584 | |||
| 585 | 1. Extract 3-5 factual claims | ||
| 586 | 2. For each claim: | ||
| 587 | - Determine verdict (WELL-SUPPORTED/PARTIALLY/UNCERTAIN/REFUTED) | ||
| 588 | - Assign confidence score (0-100%) | ||
| 589 | - Assign risk tier (A/B/C) | ||
| 590 | - Write brief reasoning (1-3 sentences) | ||
| 591 | 3. Generate analysis summary (3-5 sentences) | ||
| 592 | 4. Generate article summary (3-5 sentences) | ||
| 593 | 5. Run basic quality checks | ||
| 594 | |||
| 595 | Return as structured JSON. | ||
| 596 | {{/code}} | ||
| 597 | |||
| 598 | **Processing Time:** 10-18 seconds (estimate) | ||
| 599 | |||
| 600 | === 7.2 Full System AKEL (Production) === | ||
| 601 | |||
| 602 | **Architecture:** | ||
| 603 | {{code}} | ||
| 604 | AKEL Orchestrator | ||
| 605 | ├── Claim Extractor | ||
| 606 | ├── Claim Classifier (with risk tier assignment) | ||
| 607 | ├── Scenario Generator | ||
| 608 | ├── Evidence Summarizer | ||
| 609 | ├── Contradiction Detector | ||
| 610 | ├── Quality Gate Validator | ||
| 611 | ├── Audit Sampling Scheduler | ||
| 612 | └── Federation Sync Adapter (Release 1.0+) | ||
| 613 | {{/code}} | ||
| 614 | |||
| 615 | **Processing:** | ||
| 616 | * Parallel processing where possible | ||
| 617 | * Separate component calls | ||
| 618 | * Quality gates between phases | ||
| 619 | * Audit sampling selection | ||
| 620 | * Cross-node coordination (federated mode) | ||
| 621 | |||
| 622 | **Processing Time:** 10-30 seconds (full pipeline) | ||
| 623 | |||
| 624 | === 7.3 Why POC Uses Single Call === | ||
| 625 | |||
| 626 | **Advantages:** | ||
| 627 | * ✅ Simpler to implement | ||
| 628 | * ✅ Faster POC development | ||
| 629 | * ✅ Easier to debug | ||
| 630 | * ✅ Proves AI capability | ||
| 631 | * ✅ Good enough for concept validation | ||
| 632 | |||
| 633 | **Limitations:** | ||
| 634 | * ❌ No component reusability | ||
| 635 | * ❌ No parallel processing | ||
| 636 | * ❌ All-or-nothing (can't partially succeed) | ||
| 637 | * ❌ Harder to improve individual components | ||
| 638 | * ❌ No audit sampling | ||
| 639 | |||
| 640 | **Acceptable Trade-off:** | ||
| 641 | |||
| 642 | POC tests "Can AI do this?" not "How should we architect it?" | ||
| 643 | |||
| 644 | Full component architecture comes in Beta after POC validates concept. | ||
| 645 | |||
| 646 | === 7.4 Evolution Path === | ||
| 647 | |||
| 648 | **POC1:** Single prompt → Prove concept | ||
| 649 | **POC2:** Add scenario component → Test full pipeline | ||
| 650 | **Beta:** Multi-component AKEL → Production architecture | ||
| 651 | **Release 1.0:** Full AKEL + Federation → Scale | ||
| 652 | |||
| 653 | == 8. Functional Requirements == | ||
| 654 | |||
| 655 | === FR-POC-1: Article Input === | ||
| 656 | |||
| 657 | **Requirement:** User can submit article for analysis | ||
| 658 | |||
| 659 | **Functionality:** | ||
| 660 | * Text input field (paste article text, up to 5000 characters) | ||
| 661 | * URL input field (paste article URL) | ||
| 662 | * "Analyze" button to trigger processing | ||
| 663 | * Loading indicator during analysis | ||
| 664 | |||
| 665 | **Excluded:** | ||
| 666 | * No user authentication | ||
| 667 | * No claim history | ||
| 668 | * No search functionality | ||
| 669 | * No saved templates | ||
| 670 | |||
| 671 | **Acceptance Criteria:** | ||
| 672 | * User can paste text from article | ||
| 673 | * User can paste URL of article | ||
| 674 | * System accepts input and triggers analysis | ||
| 675 | |||
| 676 | === FR-POC-2: Claim Extraction (Fully Automated) === | ||
| 677 | |||
| 678 | **Requirement:** AI automatically extracts 3-5 factual claims | ||
| 679 | |||
| 680 | **Functionality:** | ||
| 681 | * AI reads article text | ||
| 682 | * AI identifies factual claims (not opinions/questions) | ||
| 683 | * AI extracts 3-5 most important claims | ||
| 684 | * System displays numbered list | ||
| 685 | |||
| 686 | **Critical:** NO MANUAL EDITING ALLOWED | ||
| 687 | * AI selects which claims to extract | ||
| 688 | * AI identifies factual vs. non-factual | ||
| 689 | * System processes claims as extracted | ||
| 690 | * No human curation or correction | ||
| 691 | |||
| 692 | **Error Handling:** | ||
| 693 | * If extraction fails: Display error message | ||
| 694 | * User can retry with different input | ||
| 695 | * No manual intervention to fix extraction | ||
| 696 | |||
| 697 | **Acceptance Criteria:** | ||
| 698 | * AI extracts 3-5 claims automatically | ||
| 699 | * Claims are factual (not opinions) | ||
| 700 | * Claims are clearly stated | ||
| 701 | * No manual editing required | ||
| 702 | |||
| 703 | === FR-POC-3: Verdict Generation (Fully Automated) === | ||
| 704 | |||
| 705 | **Requirement:** AI automatically generates verdict for each claim | ||
| 706 | |||
| 707 | **Functionality:** | ||
| 708 | * For each claim, AI: | ||
| 709 | * Evaluates claim based on available evidence/knowledge | ||
| 710 | * Determines verdict: WELL-SUPPORTED / PARTIALLY SUPPORTED / UNCERTAIN / REFUTED | ||
| 711 | * Assigns confidence score (0-100%) | ||
| 712 | * Assigns risk tier (A/B/C) | ||
| 713 | * Writes brief reasoning (1-3 sentences) | ||
| 714 | * System displays verdict for each claim | ||
| 715 | |||
| 716 | **Critical:** NO MANUAL EDITING ALLOWED | ||
| 717 | * AI computes verdicts based on evidence | ||
| 718 | * AI generates confidence scores | ||
| 719 | * AI writes reasoning | ||
| 720 | * No human review or adjustment | ||
| 721 | |||
| 722 | **Error Handling:** | ||
| 723 | * If verdict generation fails: Display error message | ||
| 724 | * User can retry | ||
| 725 | * No manual intervention to adjust verdicts | ||
| 726 | |||
| 727 | **Acceptance Criteria:** | ||
| 728 | * Each claim has a verdict | ||
| 729 | * Confidence score is displayed (0-100%) | ||
| 730 | * Risk tier is displayed (A/B/C) | ||
| 731 | * Reasoning is understandable (1-3 sentences) | ||
| 732 | * Verdict is defensible given reasoning | ||
| 733 | * All generated automatically by AI | ||
| 734 | |||
| 735 | === FR-POC-4: Analysis Summary (Fully Automated) === | ||
| 736 | |||
| 737 | **Requirement:** AI generates brief summary of analysis | ||
| 738 | |||
| 739 | **Functionality:** | ||
| 740 | * AI summarizes findings in 3-5 sentences: | ||
| 741 | * How many claims found | ||
| 742 | * Distribution of verdicts | ||
| 743 | * Overall assessment | ||
| 744 | * System displays at top of results | ||
| 745 | |||
| 746 | **Critical:** NO MANUAL EDITING ALLOWED | ||
| 747 | |||
| 748 | **Acceptance Criteria:** | ||
| 749 | * Summary is coherent | ||
| 750 | * Accurately reflects analysis | ||
| 751 | * 3-5 sentences | ||
| 752 | * Automatically generated | ||
| 753 | |||
| 754 | === FR-POC-5: Article Summary (Fully Automated, Optional) === | ||
| 755 | |||
| 756 | **Requirement:** AI generates brief summary of original article | ||
| 757 | |||
| 758 | **Functionality:** | ||
| 759 | * AI summarizes article content (not FactHarbor's analysis) | ||
| 760 | * 3-5 sentences | ||
| 761 | * System displays | ||
| 762 | |||
| 763 | **Note:** Optional - can skip if time limited | ||
| 764 | |||
| 765 | **Critical:** NO MANUAL EDITING ALLOWED | ||
| 766 | |||
| 767 | **Acceptance Criteria:** | ||
| 768 | * Summary is neutral (article's position) | ||
| 769 | * Accurately reflects article content | ||
| 770 | * 3-5 sentences | ||
| 771 | * Automatically generated | ||
| 772 | |||
| 773 | === FR-POC-6: Publication Mode Display === | ||
| 774 | |||
| 775 | **Requirement:** Clear labeling of AI-generated content | ||
| 776 | |||
| 777 | **Functionality:** | ||
| 778 | * Display Mode 2 publication label | ||
| 779 | * Show POC/Demo disclaimer | ||
| 780 | * Display risk tiers per claim | ||
| 781 | * Show quality gate status | ||
| 782 | * Display timestamp | ||
| 783 | |||
| 784 | **Acceptance Criteria:** | ||
| 785 | * Label is prominent and clear | ||
| 786 | * User understands this is AI-generated POC output | ||
| 787 | * Risk tiers are color-coded | ||
| 788 | * Quality gate status is visible | ||
| 789 | |||
| 790 | === FR-POC-7: Quality Gate Execution === | ||
| 791 | |||
| 792 | **Requirement:** Execute simplified quality gates | ||
| 793 | |||
| 794 | **Functionality:** | ||
| 795 | * Check source quality (basic) | ||
| 796 | * Attempt contradiction search (basic) | ||
| 797 | * Calculate confidence scores | ||
| 798 | * Verify structural integrity (basic) | ||
| 799 | * Display gate results | ||
| 800 | |||
| 801 | **Acceptance Criteria:** | ||
| 802 | * All 4 gates attempted | ||
| 803 | * Pass/fail status displayed | ||
| 804 | * Failures explained to user | ||
| 805 | * Gates don't block publication (POC mode) | ||
| 806 | |||
| 807 | == 9. Non-Functional Requirements == | ||
| 808 | |||
| 809 | === NFR-POC-1: Fully Automated Processing === | ||
| 810 | |||
| 811 | **Requirement:** Complete AI automation with zero manual intervention | ||
| 812 | |||
| 813 | **Critical Rule:** NO MANUAL EDITING AT ANY STAGE | ||
| 814 | |||
| 815 | **What this means:** | ||
| 816 | * Claims: AI selects (no human curation) | ||
| 817 | * Scenarios: N/A (deferred to POC2) | ||
| 818 | * Evidence: AI evaluates (no human selection) | ||
| 819 | * Verdicts: AI determines (no human adjustment) | ||
| 820 | * Summaries: AI writes (no human editing) | ||
| 821 | |||
| 822 | **Pipeline:** | ||
| 823 | {{code}} | ||
| 824 | User Input → AKEL Processing → Output Display | ||
| 825 | ↓ | ||
| 826 | ZERO human editing | ||
| 827 | {{/code}} | ||
| 828 | |||
| 829 | **If AI output is poor:** | ||
| 830 | * ❌ Do NOT manually fix it | ||
| 831 | * ✅ Document the failure | ||
| 832 | * ✅ Improve prompts and retry | ||
| 833 | * ✅ Accept that POC might fail | ||
| 834 | |||
| 835 | **Why this matters:** | ||
| 836 | * Tests whether AI can do this without humans | ||
| 837 | * Validates scalability (humans can't review every analysis) | ||
| 838 | * Honest test of technical feasibility | ||
| 839 | |||
| 840 | === NFR-POC-2: Performance === | ||
| 841 | |||
| 842 | **Requirement:** Analysis completes in reasonable time | ||
| 843 | |||
| 844 | **Acceptable Performance:** | ||
| 845 | * Processing time: 1-5 minutes (acceptable for POC) | ||
| 846 | * Display loading indicator to user | ||
| 847 | * Show progress if possible ("Extracting claims...", "Generating verdicts...") | ||
| 848 | |||
| 849 | **Not Required:** | ||
| 850 | * Production-level speed (< 30 seconds) | ||
| 851 | * Optimization for scale | ||
| 852 | * Caching | ||
| 853 | |||
| 854 | **Acceptance Criteria:** | ||
| 855 | * Analysis completes within 5 minutes | ||
| 856 | * User sees loading indicator | ||
| 857 | * No timeout errors | ||
| 858 | |||
| 859 | === NFR-POC-3: Reliability === | ||
| 860 | |||
| 861 | **Requirement:** System works for manual testing sessions | ||
| 862 | |||
| 863 | **Acceptable:** | ||
| 864 | * Occasional errors (< 20% failure rate) | ||
| 865 | * Manual restart if needed | ||
| 866 | * Display error messages clearly | ||
| 867 | |||
| 868 | **Not Required:** | ||
| 869 | * 99.9% uptime | ||
| 870 | * Automatic error recovery | ||
| 871 | * Production monitoring | ||
| 872 | |||
| 873 | **Acceptance Criteria:** | ||
| 874 | * System works for test demonstrations | ||
| 875 | * Errors are handled gracefully | ||
| 876 | * User receives clear error messages | ||
| 877 | |||
| 878 | === NFR-POC-4: Environment === | ||
| 879 | |||
| 880 | **Requirement:** Runs on simple infrastructure | ||
| 881 | |||
| 882 | **Acceptable:** | ||
| 883 | * Single machine or simple cloud setup | ||
| 884 | * No distributed architecture | ||
| 885 | * No load balancing | ||
| 886 | * No redundancy | ||
| 887 | * Local development environment viable | ||
| 888 | |||
| 889 | **Not Required:** | ||
| 890 | * Production infrastructure | ||
| 891 | * Multi-region deployment | ||
| 892 | * Auto-scaling | ||
| 893 | * Disaster recovery | ||
| 894 | |||
| 895 | === NFR-POC-5: Cost Efficiency Tracking === | ||
| 896 | |||
| 897 | **Requirement:** Track and display LLM usage metrics to inform optimization decisions | ||
| 898 | |||
| 899 | **Must Track:** | ||
| 900 | * Input tokens (article + prompt) | ||
| 901 | * Output tokens (generated analysis) | ||
| 902 | * Total tokens | ||
| 903 | * Estimated cost (USD) | ||
| 904 | * Response time (seconds) | ||
| 905 | * Article length (words/characters) | ||
| 906 | |||
| 907 | **Must Display:** | ||
| 908 | * Usage statistics in UI (Component 5) | ||
| 909 | * Cost per analysis | ||
| 910 | * Cost per claim extracted | ||
| 911 | |||
| 912 | **Must Log:** | ||
| 913 | * Aggregate metrics for analysis | ||
| 914 | * Cost distribution by article length | ||
| 915 | * Token efficiency trends | ||
| 916 | |||
| 917 | **Purpose:** | ||
| 918 | * Understand unit economics | ||
| 919 | * Identify optimization opportunities | ||
| 920 | * Project costs at scale | ||
| 921 | * Inform architecture decisions (caching, model selection, etc.) | ||
| 922 | |||
| 923 | **Acceptance Criteria:** | ||
| 924 | * ✅ Usage data displayed after each analysis | ||
| 925 | * ✅ Metrics logged for aggregate analysis | ||
| 926 | * ✅ Cost calculated accurately (Claude API pricing) | ||
| 927 | * ✅ Test cases include varying article lengths | ||
| 928 | * ✅ POC1 report includes cost analysis section | ||
| 929 | |||
| 930 | **Success Target:** | ||
| 931 | * Average cost per analysis < $0.05 USD | ||
| 932 | * Cost scaling behavior understood (linear/exponential) | ||
| 933 | * 2+ optimization opportunities identified | ||
| 934 | |||
| 935 | **Critical:** Unit economics must be viable for scaling decision! | ||
| 936 | |||
| 937 | == 10. Technical Architecture == | ||
| 938 | |||
| 939 | === 10.1 System Components === | ||
| 940 | |||
| 941 | **Frontend:** | ||
| 942 | * Simple HTML form (text input + URL input + button) | ||
| 943 | * Loading indicator | ||
| 944 | * Results display page (single page, no tabs/navigation) | ||
| 945 | |||
| 946 | **Backend:** | ||
| 947 | * Single API endpoint | ||
| 948 | * Calls provider API (REASONING model; configured via LLM abstraction) | ||
| 949 | * Parses response | ||
| 950 | * Returns JSON to frontend | ||
| 951 | |||
| 952 | **Data Storage:** | ||
| 953 | * None required (stateless POC) | ||
| 954 | * Optional: Simple file storage or SQLite for demo examples | ||
| 955 | |||
| 956 | **External Services:** | ||
| 957 | * Claude API (Anthropic) - required | ||
| 958 | * Optional: URL fetch service for article text extraction | ||
| 959 | |||
| 960 | === 10.2 Processing Flow === | ||
| 961 | |||
| 962 | {{code}} | ||
| 963 | 1. User submits text or URL | ||
| 964 | ↓ | ||
| 965 | 2. Backend receives request | ||
| 966 | ↓ | ||
| 967 | 3. If URL: Fetch article text | ||
| 968 | ↓ | ||
| 969 | 4. Call Claude API with single prompt: | ||
| 970 | "Extract claims, evaluate each, provide verdicts" | ||
| 971 | ↓ | ||
| 972 | 5. Claude API returns: | ||
| 973 | - Analysis summary | ||
| 974 | - Claims list | ||
| 975 | - Verdicts for each claim (with risk tiers) | ||
| 976 | - Article summary (optional) | ||
| 977 | - Quality gate results | ||
| 978 | ↓ | ||
| 979 | 6. Backend parses response | ||
| 980 | ↓ | ||
| 981 | 7. Frontend displays results with Mode 2 labeling | ||
| 982 | {{/code}} | ||
| 983 | |||
| 984 | **Key Simplification:** Single API call does entire analysis | ||
| 985 | |||
| 986 | === 10.3 AI Prompt Strategy === | ||
| 987 | |||
| 988 | **Single Comprehensive Prompt:** | ||
| 989 | {{code}} | ||
| 990 | Task: Analyze this article and provide: | ||
| 991 | |||
| 992 | 1. Identify the article's main thesis/conclusion | ||
| 993 | - What is the article trying to argue or prove? | ||
| 994 | - What is the primary claim or conclusion? | ||
| 995 | |||
| 996 | 2. Extract 3-5 factual claims from the article | ||
| 997 | - Note which claims are CENTRAL to the main thesis | ||
| 998 | - Note which claims are SUPPORTING facts | ||
| 999 | |||
| 1000 | 3. For each claim: | ||
| 1001 | - Determine verdict (WELL-SUPPORTED / PARTIALLY SUPPORTED / UNCERTAIN / REFUTED) | ||
| 1002 | - Assign confidence score (0-100%) | ||
| 1003 | - Assign risk tier (A: Medical/Legal/Safety, B: Policy/Science, C: Facts/Definitions) | ||
| 1004 | - Write brief reasoning (1-3 sentences) | ||
| 1005 | |||
| 1006 | 4. Assess relationship between claims and main thesis: | ||
| 1007 | - Do the claims actually support the article's conclusion? | ||
| 1008 | - Are there logical leaps or unsupported inferences? | ||
| 1009 | - Is the article's framing misleading even if individual facts are accurate? | ||
| 1010 | |||
| 1011 | 5. Run quality gates: | ||
| 1012 | - Check: ≥2 sources found | ||
| 1013 | - Attempt: Basic contradiction search | ||
| 1014 | - Calculate: Confidence scores | ||
| 1015 | - Verify: Structural integrity | ||
| 1016 | |||
| 1017 | 6. Write context-aware analysis summary (4-6 sentences): | ||
| 1018 | - State article's main thesis | ||
| 1019 | - Report claims found and verdict distribution | ||
| 1020 | - Note if central claims are problematic | ||
| 1021 | - Assess whether evidence supports conclusion | ||
| 1022 | - Overall credibility considering claim importance | ||
| 1023 | |||
| 1024 | 7. Write article summary (3-5 sentences: neutral summary of article content) | ||
| 1025 | |||
| 1026 | Return as structured JSON with quality gate results. | ||
| 1027 | {{/code}} | ||
| 1028 | |||
| 1029 | **One prompt generates everything.** | ||
| 1030 | |||
| 1031 | **Critical Addition:** | ||
| 1032 | |||
| 1033 | Steps 1, 2 (marking central claims), 4, and 6 are NEW for context-aware analysis. These test whether AI can distinguish between "accurate facts poorly reasoned" vs. "genuinely credible article." | ||
| 1034 | |||
| 1035 | === 10.4 Technology Stack Suggestions === | ||
| 1036 | |||
| 1037 | **Frontend:** | ||
| 1038 | * HTML + CSS + JavaScript (minimal framework) | ||
| 1039 | * OR: Next.js (if team prefers) | ||
| 1040 | * Hosted: Local machine OR Vercel/Netlify free tier | ||
| 1041 | |||
| 1042 | **Backend:** | ||
| 1043 | * Python Flask/FastAPI (simple REST API) | ||
| 1044 | * OR: Next.js API routes (if using Next.js) | ||
| 1045 | * Hosted: Local machine OR Railway/Render free tier | ||
| 1046 | |||
| 1047 | **AKEL Integration:** | ||
| 1048 | * Claude API via Anthropic SDK | ||
| 1049 | * Model: Provider-default REASONING model or latest available | ||
| 1050 | |||
| 1051 | **Database:** | ||
| 1052 | * None (stateless acceptable) | ||
| 1053 | * OR: SQLite if want to store demo examples | ||
| 1054 | * OR: JSON files on disk | ||
| 1055 | |||
| 1056 | **Deployment:** | ||
| 1057 | * Local development environment sufficient for POC | ||
| 1058 | * Optional: Deploy to cloud for remote demos | ||
| 1059 | |||
| 1060 | == 11. Success Criteria == | ||
| 1061 | |||
| 1062 | === 11.1 Minimum Success (POC Passes) === | ||
| 1063 | |||
| 1064 | **Required for GO decision:** | ||
| 1065 | * ✅ AI extracts 3-5 factual claims automatically | ||
| 1066 | * ✅ AI provides verdict for each claim automatically | ||
| 1067 | * ✅ Verdicts are reasonable (≥70% make logical sense) | ||
| 1068 | * ✅ Analysis summary is coherent | ||
| 1069 | * ✅ Output is comprehensible to reviewers | ||
| 1070 | * ✅ Team/advisors understand the output | ||
| 1071 | * ✅ Team agrees approach has merit | ||
| 1072 | * ✅ **Minimal or no manual editing needed** (< 30% of analyses require manual intervention) | ||
| 1073 | * ✅ **Cost efficiency acceptable** (average cost per analysis < $0.05 USD target) | ||
| 1074 | * ✅ **Cost scaling understood** (data collected on article length vs. cost) | ||
| 1075 | * ✅ **Optimization opportunities identified** (≥2 potential improvements documented) | ||
| 1076 | |||
| 1077 | **Quality Definition:** | ||
| 1078 | * "Reasonable verdict" = Defensible given general knowledge | ||
| 1079 | * "Coherent summary" = Logically structured, grammatically correct | ||
| 1080 | * "Comprehensible" = Reviewers understand what analysis means | ||
| 1081 | |||
| 1082 | === 11.2 POC Fails If === | ||
| 1083 | |||
| 1084 | **Automatic NO-GO if any of these:** | ||
| 1085 | * ❌ Claim extraction poor (< 60% accuracy - extracts non-claims or misses obvious ones) | ||
| 1086 | * ❌ Verdicts nonsensical (< 60% reasonable - contradictory or random) | ||
| 1087 | * ❌ Output incomprehensible (reviewers can't understand analysis) | ||
| 1088 | * ❌ **Requires manual editing for most analyses** (> 50% need human correction) | ||
| 1089 | * ❌ Team loses confidence in AI-automated approach | ||
| 1090 | |||
| 1091 | === 11.3 Quality Thresholds === | ||
| 1092 | |||
| 1093 | **POC quality expectations:** | ||
| 1094 | |||
| 1095 | |=Component|=Quality Threshold|=Definition | ||
| 1096 | |Claim Extraction|(% class="success" %)≥70% accuracy(%%) |Identifies obvious factual claims, may miss some edge cases | ||
| 1097 | |Verdict Logic|(% class="success" %)≥70% defensible(%%) |Verdicts are logical given reasoning provided | ||
| 1098 | |Reasoning Clarity|(% class="success" %)≥70% clear(%%) |1-3 sentences are understandable and relevant | ||
| 1099 | |Overall Analysis|(% class="success" %)≥70% useful(%%) |Output helps user understand article claims | ||
| 1100 | |||
| 1101 | **Analogy:** "B student" quality (70-80%), not "A+" perfection yet | ||
| 1102 | |||
| 1103 | **Not expecting:** | ||
| 1104 | * 100% accuracy | ||
| 1105 | * Perfect claim coverage | ||
| 1106 | * Comprehensive evidence gathering | ||
| 1107 | * Flawless verdicts | ||
| 1108 | * Production polish | ||
| 1109 | |||
| 1110 | **Expecting:** | ||
| 1111 | * Reasonable claim extraction | ||
| 1112 | * Defensible verdicts | ||
| 1113 | * Understandable reasoning | ||
| 1114 | * Useful output | ||
| 1115 | |||
| 1116 | == 12. Test Cases == | ||
| 1117 | |||
| 1118 | === 12.1 Test Case 1: Simple Factual Claim === | ||
| 1119 | |||
| 1120 | **Input:** "Coffee reduces the risk of type 2 diabetes by 30%" | ||
| 1121 | |||
| 1122 | **Expected Output:** | ||
| 1123 | * Extract claim correctly | ||
| 1124 | * Provide verdict: WELL-SUPPORTED or PARTIALLY SUPPORTED | ||
| 1125 | * Confidence: 70-90% | ||
| 1126 | * Risk tier: C (Low) | ||
| 1127 | * Reasoning: Mentions studies or evidence | ||
| 1128 | |||
| 1129 | **Success:** Verdict is reasonable and reasoning makes sense | ||
| 1130 | |||
| 1131 | === 12.2 Test Case 2: Complex News Article === | ||
| 1132 | |||
| 1133 | **Input:** News article URL with multiple claims about politics/health/science | ||
| 1134 | |||
| 1135 | **Expected Output:** | ||
| 1136 | * Extract 3-5 key claims | ||
| 1137 | * Verdict for each (may vary: some supported, some uncertain, some refuted) | ||
| 1138 | * Coherent analysis summary | ||
| 1139 | * Article summary | ||
| 1140 | * Risk tiers assigned appropriately | ||
| 1141 | |||
| 1142 | **Success:** Claims identified are actually from article, verdicts are reasonable | ||
| 1143 | |||
| 1144 | === 12.3 Test Case 3: Controversial Topic === | ||
| 1145 | |||
| 1146 | **Input:** Article on contested political or scientific topic | ||
| 1147 | |||
| 1148 | **Expected Output:** | ||
| 1149 | * Balanced analysis | ||
| 1150 | * Acknowledges uncertainty where appropriate | ||
| 1151 | * Doesn't overstate confidence | ||
| 1152 | * Reasoning shows awareness of complexity | ||
| 1153 | |||
| 1154 | **Success:** Analysis is fair and doesn't show obvious bias | ||
| 1155 | |||
| 1156 | === 12.4 Test Case 4: Clearly False Claim === | ||
| 1157 | |||
| 1158 | **Input:** Article with obviously false claim (e.g., "The Earth is flat") | ||
| 1159 | |||
| 1160 | **Expected Output:** | ||
| 1161 | * Extract claim | ||
| 1162 | * Verdict: REFUTED | ||
| 1163 | * High confidence (> 90%) | ||
| 1164 | * Risk tier: C (Low - established fact) | ||
| 1165 | * Clear reasoning | ||
| 1166 | |||
| 1167 | **Success:** AI correctly identifies false claim with high confidence | ||
| 1168 | |||
| 1169 | === 12.5 Test Case 5: Genuinely Uncertain Claim === | ||
| 1170 | |||
| 1171 | **Input:** Article with claim where evidence is genuinely mixed | ||
| 1172 | |||
| 1173 | **Expected Output:** | ||
| 1174 | * Extract claim | ||
| 1175 | * Verdict: UNCERTAIN | ||
| 1176 | * Moderate confidence (40-60%) | ||
| 1177 | * Reasoning explains why uncertain | ||
| 1178 | |||
| 1179 | **Success:** AI recognizes uncertainty and doesn't overstate confidence | ||
| 1180 | |||
| 1181 | === 12.6 Test Case 6: High-Risk Medical Claim === | ||
| 1182 | |||
| 1183 | **Input:** Article making medical claims | ||
| 1184 | |||
| 1185 | **Expected Output:** | ||
| 1186 | * Extract claim | ||
| 1187 | * Verdict: [appropriate based on evidence] | ||
| 1188 | * Risk tier: A (High - medical) | ||
| 1189 | * Red label displayed | ||
| 1190 | * Clear disclaimer about not being medical advice | ||
| 1191 | |||
| 1192 | **Success:** Risk tier correctly assigned, appropriate warnings shown | ||
| 1193 | |||
| 1194 | == 13. POC Decision Gate == | ||
| 1195 | |||
| 1196 | === 13.1 Decision Framework === | ||
| 1197 | |||
| 1198 | After POC testing complete, team makes one of three decisions: | ||
| 1199 | |||
| 1200 | **Option A: GO (Proceed to POC2)** | ||
| 1201 | |||
| 1202 | **Conditions:** | ||
| 1203 | * AI quality ≥70% without manual editing | ||
| 1204 | * Basic claim → verdict pipeline validated | ||
| 1205 | * Internal + advisor feedback positive | ||
| 1206 | * Technical feasibility confirmed | ||
| 1207 | * Team confident in direction | ||
| 1208 | * Clear path to improving AI quality to ≥90% | ||
| 1209 | |||
| 1210 | **Next Steps:** | ||
| 1211 | * Plan POC2 development (add scenarios) | ||
| 1212 | * Design scenario architecture | ||
| 1213 | * Expand to Evidence Model structure | ||
| 1214 | * Test with more complex articles | ||
| 1215 | |||
| 1216 | **Option B: NO-GO (Pivot or Stop)** | ||
| 1217 | |||
| 1218 | **Conditions:** | ||
| 1219 | * AI quality < 60% | ||
| 1220 | * Requires manual editing for most analyses (> 50%) | ||
| 1221 | * Feedback indicates fundamental flaws | ||
| 1222 | * Cost/effort not justified by value | ||
| 1223 | * No clear path to improvement | ||
| 1224 | |||
| 1225 | **Next Steps:** | ||
| 1226 | * **Pivot:** Change to hybrid human-AI approach (accept manual review required) | ||
| 1227 | * **Stop:** Conclude approach not viable, revisit later | ||
| 1228 | |||
| 1229 | **Option C: ITERATE (Improve POC)** | ||
| 1230 | |||
| 1231 | **Conditions:** | ||
| 1232 | * Concept has merit but execution needs work | ||
| 1233 | * Specific improvements identified | ||
| 1234 | * Addressable with better prompts/approach | ||
| 1235 | * AI quality between 60-70% | ||
| 1236 | |||
| 1237 | **Next Steps:** | ||
| 1238 | * Improve AI prompts | ||
| 1239 | * Test different approaches | ||
| 1240 | * Re-run POC with improvements | ||
| 1241 | * Then make GO/NO-GO decision | ||
| 1242 | |||
| 1243 | === 13.2 Decision Criteria Summary === | ||
| 1244 | |||
| 1245 | {{code}} | ||
| 1246 | AI Quality < 60% → NO-GO (approach doesn't work) | ||
| 1247 | AI Quality 60-70% → ITERATE (improve and retry) | ||
| 1248 | AI Quality ≥70% → GO (proceed to POC2) | ||
| 1249 | {{/code}} | ||
| 1250 | |||
| 1251 | == 14. Key Risks & Mitigations == | ||
| 1252 | |||
| 1253 | === 14.1 Risk: AI Quality Not Good Enough === | ||
| 1254 | |||
| 1255 | **Likelihood:** Medium-High | ||
| 1256 | **Impact:** POC fails | ||
| 1257 | |||
| 1258 | **Mitigation:** | ||
| 1259 | * Extensive prompt engineering and testing | ||
| 1260 | * Use best available AI models (role-based selection; configured via LLM abstraction) | ||
| 1261 | * Test with diverse article types | ||
| 1262 | * Iterate on prompts based on results | ||
| 1263 | |||
| 1264 | **Acceptance:** This is what POC tests - be ready for failure | ||
| 1265 | |||
| 1266 | === 14.2 Risk: AI Consistency Issues === | ||
| 1267 | |||
| 1268 | **Likelihood:** Medium | ||
| 1269 | **Impact:** Works sometimes, fails other times | ||
| 1270 | |||
| 1271 | **Mitigation:** | ||
| 1272 | * Test with 10+ diverse articles | ||
| 1273 | * Measure success rate honestly | ||
| 1274 | * Improve prompts to increase consistency | ||
| 1275 | |||
| 1276 | **Acceptance:** Some variability OK if average quality ≥70% | ||
| 1277 | |||
| 1278 | === 14.3 Risk: Output Incomprehensible === | ||
| 1279 | |||
| 1280 | **Likelihood:** Low-Medium | ||
| 1281 | **Impact:** Users can't understand analysis | ||
| 1282 | |||
| 1283 | **Mitigation:** | ||
| 1284 | * Create clear explainer document | ||
| 1285 | * Iterate on output format | ||
| 1286 | * Test with non-technical reviewers | ||
| 1287 | * Simplify language if needed | ||
| 1288 | |||
| 1289 | **Acceptance:** Iterate until comprehensible | ||
| 1290 | |||
| 1291 | === 14.4 Risk: API Rate Limits / Costs === | ||
| 1292 | |||
| 1293 | **Likelihood:** Low | ||
| 1294 | **Impact:** System slow or expensive | ||
| 1295 | |||
| 1296 | **Mitigation:** | ||
| 1297 | * Monitor API usage | ||
| 1298 | * Implement retry logic | ||
| 1299 | * Estimate costs before scaling | ||
| 1300 | |||
| 1301 | **Acceptance:** POC can be slow and expensive (optimization later) | ||
| 1302 | |||
| 1303 | === 14.5 Risk: Scope Creep === | ||
| 1304 | |||
| 1305 | **Likelihood:** Medium | ||
| 1306 | **Impact:** POC becomes too complex | ||
| 1307 | |||
| 1308 | **Mitigation:** | ||
| 1309 | * Strict scope discipline | ||
| 1310 | * Say NO to feature additions | ||
| 1311 | * Keep focus on core question | ||
| 1312 | |||
| 1313 | **Acceptance:** POC is minimal by design | ||
| 1314 | |||
| 1315 | == 15. POC Philosophy == | ||
| 1316 | |||
| 1317 | === 15.1 Core Principles === | ||
| 1318 | |||
| 1319 | **1. Build Less, Learn More** | ||
| 1320 | * Minimum features to test hypothesis | ||
| 1321 | * Don't build unvalidated features | ||
| 1322 | * Focus on core question only | ||
| 1323 | |||
| 1324 | **2. Fail Fast** | ||
| 1325 | * Quick test of hardest part (AI capability) | ||
| 1326 | * Accept that POC might fail | ||
| 1327 | * Better to discover issues early | ||
| 1328 | * Honest assessment over optimistic hope | ||
| 1329 | |||
| 1330 | **3. Test First, Build Second** | ||
| 1331 | * Validate AI can do this before building platform | ||
| 1332 | * Don't assume it will work | ||
| 1333 | * Let results guide decisions | ||
| 1334 | |||
| 1335 | **4. Automation First** | ||
| 1336 | * No manual editing allowed | ||
| 1337 | * Tests scalability, not just feasibility | ||
| 1338 | * Proves approach can work at scale | ||
| 1339 | |||
| 1340 | **5. Honest Assessment** | ||
| 1341 | * Don't cherry-pick examples | ||
| 1342 | * Don't manually fix bad outputs | ||
| 1343 | * Document failures openly | ||
| 1344 | * Make data-driven decisions | ||
| 1345 | |||
| 1346 | === 15.2 What POC Is === | ||
| 1347 | |||
| 1348 | ✅ Testing AI capability without humans | ||
| 1349 | ✅ Proving core technical concept | ||
| 1350 | ✅ Fast validation of approach | ||
| 1351 | ✅ Honest assessment of feasibility | ||
| 1352 | |||
| 1353 | === 15.3 What POC Is NOT === | ||
| 1354 | |||
| 1355 | ❌ Building a product | ||
| 1356 | ❌ Production-ready system | ||
| 1357 | ❌ Feature-complete platform | ||
| 1358 | ❌ Perfectly accurate analysis | ||
| 1359 | ❌ Polished user experience | ||
| 1360 | |||
| 1361 | == 16. Success = Clear Path Forward == | ||
| 1362 | |||
| 1363 | **If POC succeeds (≥70% AI quality):** | ||
| 1364 | * ✅ Approach validated | ||
| 1365 | * ✅ Proceed to POC2 (add scenarios) | ||
| 1366 | * ✅ Design full Evidence Model structure | ||
| 1367 | * ✅ Test multi-scenario comparison | ||
| 1368 | * ✅ Focus on improving AI quality from 70% → 90% | ||
| 1369 | |||
| 1370 | **If POC fails (< 60% AI quality):** | ||
| 1371 | * ✅ Learn what doesn't work | ||
| 1372 | * ✅ Pivot to different approach | ||
| 1373 | * ✅ OR wait for better AI technology | ||
| 1374 | * ✅ Avoid wasting resources on non-viable approach | ||
| 1375 | |||
| 1376 | **Either way, POC provides clarity.** | ||
| 1377 | |||
| 1378 | == 17. Related Pages == | ||
| 1379 | |||
| 1380 | * [[User Needs>>FactHarbor.Specification.Requirements.User Needs.WebHome]] | ||
| 1381 | * [[Requirements>>FactHarbor.Specification.Requirements.WebHome]] | ||
| 1382 | * [[Gap Analysis>>FactHarbor.Specification.Requirements.GapAnalysis]] | ||
| 1383 | * [[Architecture>>FactHarbor.Specification.Architecture.WebHome]] | ||
| 1384 | * [[AKEL>>FactHarbor.Specification.AI Knowledge Extraction Layer (AKEL).WebHome]] | ||
| 1385 | * [[Workflows>>FactHarbor.Specification.Workflows.WebHome]] | ||
| 1386 | |||
| 1387 | **Document Status:** ✅ Ready for POC Development (Version 2.0 - Updated with Spec Alignment) | ||
| 1388 | |||
| 1389 | |||
| 1390 | === NFR-POC-11: LLM Provider Abstraction (POC1) === | ||
| 1391 | |||
| 1392 | **Requirement:** POC1 MUST implement LLM abstraction layer with support for multiple providers. | ||
| 1393 | |||
| 1394 | **POC1 Implementation:** | ||
| 1395 | |||
| 1396 | * **Primary Provider:** Anthropic Claude API | ||
| 1397 | * Stage 1: Provider-default FAST model | ||
| 1398 | * Stage 2: Provider-default REASONING model (cached) | ||
| 1399 | * Stage 3: Provider-default REASONING model | ||
| 1400 | |||
| 1401 | * **Provider Interface:** Abstract LLMProvider interface implemented | ||
| 1402 | |||
| 1403 | * **Configuration:** Environment variables for provider selection | ||
| 1404 | * {{code}}LLM_PRIMARY_PROVIDER=anthropic{{/code}} | ||
| 1405 | * {{code}}LLM_STAGE1_MODEL=claude-haiku-4{{/code}} | ||
| 1406 | * {{code}}LLM_STAGE2_MODEL=claude-sonnet-3-5{{/code}} | ||
| 1407 | |||
| 1408 | * **Failover:** Basic error handling with cache fallback for Stage 2 | ||
| 1409 | |||
| 1410 | * **Cost Tracking:** Log provider name and cost per request | ||
| 1411 | |||
| 1412 | **Future (POC2/Beta):** | ||
| 1413 | |||
| 1414 | * Secondary provider (OpenAI) with automatic failover | ||
| 1415 | * Admin API for runtime provider switching | ||
| 1416 | * Cost comparison dashboard | ||
| 1417 | * Cross-provider output verification | ||
| 1418 | |||
| 1419 | **Success Criteria:** | ||
| 1420 | |||
| 1421 | * All LLM calls go through abstraction layer (no direct API calls) | ||
| 1422 | * Provider can be changed via environment variable without code changes | ||
| 1423 | * Cost tracking includes provider name in logs | ||
| 1424 | * Stage 2 falls back to cache on provider failure | ||
| 1425 | |||
| 1426 | **Implementation:** See [[POC1 API & Schemas Specification>>Test.FactHarbor.Specification.POC.API-and-Schemas.WebHome]] Section 6 | ||
| 1427 | |||
| 1428 | **Dependencies:** | ||
| 1429 | * NFR-14 (Main Requirements) | ||
| 1430 | * Design Decision 9 | ||
| 1431 | * Architecture Section 2.2 | ||
| 1432 | |||
| 1433 | **Priority:** HIGH (P1) | ||
| 1434 | |||
| 1435 | **Rationale:** Even though POC1 uses single provider, abstraction must be in place from start to avoid costly refactoring later. |