# Round 13 โ€” Tower J10-J12 ยท Vector / FactCheck / Tone **Date**: 2026-04-30 **Time**: 1.5h budget **Status**: โœ… COMPLETE --- ## ๐ŸŽฏ ื”ืงื•ืžื•ืช | # | Station | ืชืคืงื™ื“ | Latency | |---|---|---|---| | **J10** | Vector Context Search | pgvector ื—ื™ืคื•ืฉ ื›ืชื‘ื•ืช ื“ื•ืžื•ืช | <800ms | | **J11** | Fact-Check Layer | ืื™ืžื•ืช ืขื•ื‘ื“ื•ืช ืžื•ืœ archive + web | <2s | | **J12** | Tone Analyzer | ื–ื™ื”ื•ื™ tone ืฉืœ ื”ื•ื“ืขืช ื”ืžืงื•ืจ (formal/breaking/promotional) | <500ms | --- ## ๐Ÿ” J10 ยท Vector Context Search ### ืชืคืงื™ื“ ืžืฆื ื›ืชื‘ื•ืช historical ืฉื“ื•ืžื•ืช ืœื˜ืงืกื˜ ื”ื ื•ื›ื—ื™. ื”-Writer ื™ืฉืชืžืฉ ื‘ื”ืŸ ืœ-context, citations, ื•-related links. ### Implementation โ€” pgvector ืžืงื•ืจ verified: `media_platform.jason_items` table (31,590 articles, vector_5w 1536-dim). ```sql -- Two-stage retrieval -- Stage 1: vector similarity (top-50) WITH similar AS ( SELECT id, title, slug, doc_date, vector_5w <=> $1::vector AS distance FROM jason_items WHERE source = ANY($2::text[]) -- maariv, walla, i24 AND doc_date > NOW() - INTERVAL '2 years' ORDER BY distance LIMIT 50 ) -- Stage 2: re-rank by metadata + freshness SELECT s.*, (1 - s.distance) * 0.7 + (1.0 / (1 + EXTRACT(DAYS FROM NOW() - s.doc_date) / 365)) * 0.3 AS combined_score FROM similar s ORDER BY combined_score DESC LIMIT 12; ``` ### Two-stage scoring rationale - **Stage 1 (cosine)**: similarity ืœืชื•ื›ืŸ (text-level) - **Stage 2 (re-rank)**: ืžื•ืกื™ืฃ **freshness boost** โ€” ื›ืชื‘ื•ืช ืžื”ืฉื‘ื•ืข ื”ืื—ืจื•ืŸ ืžืงื‘ืœื•ืช ืฆื™ื•ืŸ ื’ื‘ื•ื” ื™ื•ืชืจ. ืœืžื”? ื›ื™ **breaking news context** ืจืœื•ื•ื ื˜ื™ ื™ื•ืชืจ. ### A2UI streaming ืœื›ืœ context_match โ†’ updateComponents: ```json {"version":"v0.9","updateComponents":{"surfaceId":"...","components":[ {"id":"ctx-1","component":"ContextCard","mode":"compact","title":"ื›ื™ืคืช ื‘ืจื–ืœ - ื”ื™ืกื˜ื•ืจื™ื”","score":0.92,"deskColor":"security","image":"img/maariv/articles/2024/iron-dome.jpg","action":{"name":"context_select","context":{"id":"ctx-1"}}} ]}} ``` ื”-context grid ืžืชืžืœื ื‘-12 cards ืชื•ืš ~800ms โ€” ื—ื•ื•ื™ื™ืช "ChatGPT thinking" (cards ืžื•ืคื™ืขื™ื ืื—ื“ ืื—ื“, ืœื ื›ื•ืœื ื‘ื‘ืช ืื—ืช). ### Cost - Vector embedding (5w concat): ืฉื›ื‘ืจ ื™ื™ืฆืจื• ื‘-J05-J09 โ†’ 0 - pgvector query: ~50ms ยท ~$0 - Re-rank PostgreSQL: ~30ms ยท ~$0 - **Total cost: ~$0** (compute on EX63 GPU server) --- ## โœ… J11 ยท Fact-Check Layer ### ืชืคืงื™ื“ ืœื›ืœ **factual claim** ื‘ื˜ืงืกื˜, ื‘ื“ื•ืง: 1. **Internal**: ื”ืื ื“ื•ืžื” ื‘ืžืื’ืจ archive? (ื™ืฆื™ื‘ื•ืช / ืกืชื™ืจื”) 2. **External**: web search (ืื•ืคืฆื™ื•ื ืœื™, BBC/NYT level) 3. **Numerical**: ืื ืžืกืคืจื™ื โ€” verify (12 ืกื•ืœืœื•ืช = ืœ ฬŒ2 ืžื˜ื•ืกื™ื?) 4. **Named entities**: ืื ืฉื โ†’ exists in Wikipedia/Wikidata? ### Implementation **3-stage cascade**: #### Stage 1: Claim extraction ```python prompt = """Extract all factual claims from this text. Output JSON array, each item: { "claim": "...", "type": "numerical" | "named" | "event" | "statement", "subject": "...", "needs_verification": true | false } Reject opinions, predictions.""" ``` #### Stage 2: Per-claim verification ```python for claim in claims: if claim.type == "named": result = await wikidata_lookup(claim.subject) elif claim.type == "numerical": result = await archive_compare(claim) # search vector for similar numbers elif claim.type == "event": result = await archive_event_check(claim) # โ†’ result: { verdict: "verified" | "disputed" | "unknown", evidence: [...] } ``` #### Stage 3: Aggregate flags ```python flags = [] for claim, result in claim_results: if result.verdict == "disputed": flags.append({ "claim": claim.claim, "verdict": "disputed", "evidence": result.evidence, "severity": "high" }) yield _emit("factcheck", {"flags": flags, "total_claims": len(claims), "verified_count": ...}) ``` ### A2UI emission ```json {"version":"v0.9","updateComponents":{"surfaceId":"article-output","components":[ {"id":"fc-3","component":"FactCheckFlag","claim":"12 ืกื•ืœืœื•ืช","verdict":"disputed","evidence":"ืžืืžืจื™ื ืงื•ื“ืžื™ื ืžืฆื™ื™ื ื™ื 9 ืกื•ืœืœื•ืช","severity":"high","placement":{"line":4,"col_start":12,"col_end":24}} ]}} ``` ื”flag ืžื•ืคื™ืข **inline** ื‘ืžืืžืจ โ€” ืฆื‘ืข ืื“ื•ื ืžืกื‘ื™ื‘ ืœ-"12 ืกื•ืœืœื•ืช" + tooltip ืขื evidence. ### Cost - Claim extraction: ~$0.0003 - Per-claim verification (avg 5 claims ยท 50% need LLM): ~$0.0008 - **Total per request: ~$0.0011** --- ## ๐ŸŽญ J12 ยท Tone Analyzer ### ืชืคืงื™ื“ ื–ื™ื”ื•ื™ tone ืฉืœ ื”**ืžืงื•ืจ** (ืœื ืฉืœ ื”-output). ืœืžื”? - **Promotional press release** โ†’ ื”ื›ื•ืชื‘ ืฆืจื™ืš **ืœื ื˜ืจืœ** ืืช ื”bias ื•ืœื”ืคื•ืš ืœ-news neutral - **Breaking news from agency** โ†’ tone ื“ื—ื•ืฃ โ€” ื”ื›ื•ืชื‘ ืฆืจื™ืš ืคืฉื˜ื•ืช + ืžื”ื™ืจื•ืช - **Government statement** โ†’ tone formal โ€” ื”ื›ื•ืชื‘ ืฆืจื™ืš ืœืฉืžื•ืจ ืขืœ ื“ืจื’ืช ืจืฉืžื™ื•ืช ### Output schema ```json { "primary_tone": "promotional", "tone_scores": { "formal": 0.45, "casual": 0.05, "promotional": 0.78, "alarmist": 0.12, "academic": 0.08, "journalistic": 0.32 }, "register": "high", "neutralization_recommended": true, "writer_instructions": "Convert promotional language to neutral reporting. Strip superlatives. Remove direct quotes from PR-speak." } ``` ื”-`writer_instructions` ื ื›ืœืœ ื‘-system prompt ืฉืœ J13-J16 (Writers). ื–ื” ื”reasoning ืฉืœ ื”-Tone โ€” ืœื ืจืง ื–ื™ื”ื•ื™, ืืœื ื”ื•ืจืื•ืช ื”ืคืขื•ืœื”. ### Implementation **Single LLM call** (Gemini 2.5 Flash) ืขื few-shot examples ืฉืœ 6 tones. ### A2UI emission ```json {"version":"v0.9","updateComponents":{"surfaceId":"...","components":[ {"id":"tone-indicator","component":"BrandVoiceIndicator","sourceTone":"promotional","targetTone":"journalistic","gapScore":0.46,"warning":true} ]}} ``` ื”-`BrandVoiceIndicator` ืžืฆื™ื’ gauge โ€” ืื™ืš ื”-source tone ืžืชืื™ื ืœ-tenant brand voice (Maariv = journalistic, BBC = formal, Tikun Olam = activist). ### Cost - 1 Gemini Flash call: ~$0.0002 --- ## ๐Ÿ“Š J10-J12 ยท Total | Station | Latency | Cost | Notes | |---|---|---|---| | J10 Vector | <800ms | ~$0 | self-hosted | | J11 FactCheck | <2s | $0.0011 | most expensive | | J12 Tone | <500ms | $0.0002 | | | **Sum** | **<3.3s** | **$0.0013** | | ื”claim ื”ื•ื ืฉ-J10 ื•-J12 ืจืฆื™ื **ื‘ืžืงื‘ื™ืœ** ืขื 5W (J05-J09 ื‘-Round 12). J11 ืจืฅ **ืื—ืจื™** 5W ื›ื™ ื”ื•ื ืฆืจื™ืš claims structured. ``` [J04 Desk] โ”€โ”ฌโ”€โ†’ [J05-J09 5W] โ”€โ” โ”œโ”€โ†’ [J10 Vector] โ”œโ”€โ†’ [J11 FactCheck] โ†’ [J13 Writer] โ””โ”€โ†’ [J12 Tone] โ”€โ”˜ ``` --- ## ๐ŸŽฏ ื”ื”ื›ืจืขื•ืช ื”ืงืจื™ื˜ื™ื•ืช ### 1. ืœืžื” J11 (FactCheck) ืจืง ืื—ืจื™ 5W? ื”claims ืžื•ื‘ื ื™ื ื™ื•ืชืจ ืื—ืจื™ ืฉื™ืฉ 5W structured. ืœืžืฉืœ "12 ืกื•ืœืœื•ืช" โ€” ื”-J06 (What) ื›ื‘ืจ ื—ื™ืœืฅ "object: 12 ืกื•ืœืœื•ืช". J11 ืœื ืฆืจื™ืš ืœื—ืคืฉ ืืช ื–ื” ืฉื•ื‘. ### 2. ืœืžื” J10 (Vector) ืจืฅ ื‘ืžืงื‘ื™ืœ ืœ-5W? ื”embedding ืœื—ื™ืคื•ืฉ vector ืœื ื—ื™ื™ื‘ ืœื—ื›ื•ืช ืœ-5W. ื”ื•ื ื™ื›ื•ืœ ืœื”ืฉืชืžืฉ ื‘vector ืฉืœ ื”text ื”ืžืงื•ืจื™ (raw). ืื 5W ืžืฆื ืžืฉื”ื• ืžืขื ื™ื™ืŸ โ€” ืฉืœื‘ second pass ืขื 5w_concat embedding. ### 3. ืœืžื” J12 (Tone) ื•ืœื tenant_voice? J12 ืžื ืชื— ืืช ื”**source**, ืœื ืืช ื”**target**. ื”-target (tenant brand voice) ื”ื•ื L3 token override + writer prompt template. ืฉื ื™ ืฉืœื‘ื™ื ื ืคืจื“ื™ื. --- ## โœ… Closure โœ… **Round 13 closed. 12/54 floors documented.** --- ## ๐Ÿ›ฃ๏ธ Next: Round 14 โ€” J13-J17 (4 Writers + Editor)