# Round 11 โ€” Tower Stations J01-J04 ยท Reception Quartet **Date**: 2026-04-30 **Time**: 1.5h budget **Status**: โœ… COMPLETE โ€” first 4 floors specified **Round Type**: Tower Floors (B series begins) --- ## ๐ŸŽฏ ื”ืงื•ืžื•ืช ื‘ื”ืŸ ืขื•ืกืง ืกื‘ื‘ ื–ื” ื”ืžื’ื“ืœ ืžืชื—ื™ืœ ื‘ื›ื ื™ืกื”. ื”ื”ื•ื“ืขื” ืžื’ื™ืขื” ืืœ **ื”ืœื•ื‘ื™**, ื•ืขื•ื‘ืจืช **4 ืชื—ื ื•ืช ืงืœื™ื˜ื”** ืœืคื ื™ ืฉืžื•ืขื‘ืจืช ืœ-floors ื”ื’ื‘ื•ื”ื™ื ื™ื•ืชืจ. ื›ืœ ืชื—ื ื” ื”ื™ื LLM agent ื™ื™ืขื•ื“ื™ + cache + fallback + observability. | # | Station | ืชืคืงื™ื“ | Latency target | |---|---|---|---| | **J01** | Reception | ืงื•ืœื˜ raw input ยท validates ยท creates flow_id | <50ms | | **J02** | Language Detector | ื–ื™ื”ื•ื™ ืฉืคื” (he/en/ar) ยท confidence | <200ms | | **J03** | Triage | ืจืžืช ื“ื—ื™ืคื•ืช ยท breaking? ยท embargo? | <500ms | | **J04** | Desk Router | ืื™ื–ื” desk (security/politics/economy/...) | <800ms | **Cumulative latency** ืขื“ ืฉื”ืžืกืžืš ืขื•ื‘ืจ ืœ-J05: ~1.5 sec. ืœืื—ืจ ื–ื” ืžืชื—ื™ืœ 5W ื‘-parallel. --- ## ๐Ÿ›๏ธ J01 ยท Reception (Lobby Floor) ### ืชืคืงื™ื“ ื”-**Lobby** ืฉืœ ื”ืžื’ื“ืœ. ื›ืœ hodaแบก ืžื’ื™ืขื” ืœื›ืืŸ ืจืืฉื•ื ื”. ืชืคืงื™ื“ื•: 1. **Validate**: text > 10 chars ยท text < 100K chars ยท UTF-8 valid 2. **Sanitize**: strip HTML ยท normalize whitespace ยท detect script injection 3. **Flow creation**: INSERT into `press_flows` table โ†’ `flow_id` 4. **Tenant scoping**: extract `X-Tenant-ID` header โ†’ set RLS context 5. **Audit**: log incoming request to `press_flow_events` (event=`reception_received`) ### Implementation **ืœื LLM**. ื–ื” layer ื“ื˜ืจืžื™ื ื™ืกื˜ื™ ื‘ืœื‘ื“ โ€” Python validation + DB INSERT. ```python # /opt/maariv-press/api/routes/press_v3.py @router.post("/press/v3/stream") async def press_stream(request: Request, body: PressInput): # J01 โ€” Reception if len(body.text) < 10: raise HTTPException(400, "Text too short") if len(body.text) > 100_000: raise HTTPException(400, "Text too long (max 100K chars)") # Sanitize clean_text = bleach.clean(body.text, tags=[], strip=True) clean_text = re.sub(r'\s+', ' ', clean_text).strip() # Tenant scoping tenant_id = int(request.headers.get('X-Tenant-ID', '0')) if not tenant_id: raise HTTPException(401, "Missing tenant") # Flow creation flow_id = await db.insert_press_flow(tenant_id=tenant_id, source_text=clean_text) # Audit yield _emit("reception_received", {"flow_id": flow_id, "char_count": len(clean_text)}) # Continue to J02... ``` ### A2UI emission ```json {"version":"v0.9","createSurface":{"surfaceId":"press-flow-{flow_id}","catalogId":"...","theme":{...}}} {"version":"v0.9","updateDataModel":{"surfaceId":"...","path":"/meta/state","value":"received"}} {"version":"v0.9","updateDataModel":{"surfaceId":"...","path":"/input/length","value":1247}} ``` ### Cost - 0 LLM tokens - 0 vector queries - 1 DB INSERT - **Cost: ~$0.000001 per request** --- ## ๐ŸŒ J02 ยท Language Detector ### ืชืคืงื™ื“ ื–ื™ื”ื•ื™ ืฉืคืช ื”ืžืกืžืš ื•ื›ื™ื•ื•ืŸ (LTR/RTL). ื—ืฉื•ื‘ ื›ื™: - ืžืกืžื›ื™ื ื‘ืขื‘ืจื™ืช/ืขืจื‘ื™ืช = RTL โ†’ A2UI theme.rtl=true - BBC tenant ืžืงื‘ืœ ืื ื’ืœื™ืช โ†’ ืื ืžืงื‘ืœื™ื ืขื‘ืจื™ืช = error/warning - ื‘ื•ื—ืจ ืืช ื”-LLM prompts ื”ืžืชืื™ืžื™ื (Hebrew vs English vs Arabic) ### Implementation **ืฉื ื™ ืฉืœื‘ื™ื**: #### Stage 1: Heuristic (95% ืžื”ืžืงืจื™ื, free) ```python def detect_lang_heuristic(text: str) -> tuple[str, float]: """Fast char-set detection ยท returns (lang, confidence)""" hebrew_chars = len(re.findall(r'[\u0590-\u05FF]', text)) arabic_chars = len(re.findall(r'[\u0600-\u06FF]', text)) latin_chars = len(re.findall(r'[a-zA-Z]', text)) total = hebrew_chars + arabic_chars + latin_chars if total == 0: return ("unknown", 0.0) if hebrew_chars / total > 0.6: return ("he", hebrew_chars / total) if arabic_chars / total > 0.6: return ("ar", arabic_chars / total) if latin_chars / total > 0.7: return ("en", latin_chars / total) return ("mixed", 0.0) ``` #### Stage 2: LLM fallback (5% โ€” mixed content) ืื confidence < 0.6 โ†’ ืงืจื™ืื” ืœ-Gemini Flash: ``` prompt: "What is the primary language of this text? Reply with ISO-639-1 code only.\n\nText:\n{text}" expected: "he" | "en" | "ar" | "fr" | ... cost: ~$0.0001 ``` ### A2UI emission ```json {"version":"v0.9","updateDataModel":{"surfaceId":"...","path":"/input/language","value":"he"}} {"version":"v0.9","updateDataModel":{"surfaceId":"...","path":"/input/direction","value":"rtl"}} {"version":"v0.9","updateDataModel":{"surfaceId":"...","path":"/input/lang_confidence","value":0.97}} ``` ### Cost - Heuristic: 0 cost - LLM fallback (5% of cases): ~$0.0001 each - **Average per request: ~$0.000005** --- ## ๐Ÿšจ J03 ยท Triage (Urgency + Embargo Detection) ### ืชืคืงื™ื“ ืฉืœื™ืคืช signals ืฉืœ ื“ื—ื™ืคื•ืช ื•sensitivity: - **Urgency**: breaking / urgent / standard / archive - **Embargo**: ื”ืื ื”ืžืกืžืš ืžืกื•ืžืŸ "ื—ืกื•ื™ ืขื“ ืฉืขื” X"? - **Sensitivity**: military_security / personal_info / minors / minor_victims ื–ื” **ื—ื•ืกื ืฉืœื‘**: ืื sensitivity = `personal_info_unmasked` โ†’ flag ืœืขื•ืจืš ืœื‘ื“ื™ืงื” ืื ื•ืฉื™ืช. ### Implementation LLM Agent + rule-based hybrid: ```python SYSTEM_PROMPT = """You are a press desk triage agent for Israeli media. Analyze the press release and output JSON: { "urgency": "breaking" | "urgent" | "standard" | "archive", "urgency_reasons": [list of strings], "embargo": null | { "until": ISO_TIMESTAMP, "source": "explicit_text" }, "sensitivity": { "military": 0.0-1.0, "personal_info": 0.0-1.0, "minors": 0.0-1.0, "victims": 0.0-1.0 }, "flags": [list of warning strings] } Rules: - If text mentions IDF military operations โ†’ military: 0.8+ - If names + ages of children โ†’ minors: 0.9+ - If "embargoed until X" or "ืœื”ืคืฆื” ืจืง ืœืื—ืจ Y" โ†’ embargo set - If "killed", "wounded" + names โ†’ victims: 0.7+ """ # Gemini 2.5 Flash with structured output response = await gemini.generate( system=SYSTEM_PROMPT, user=text, response_format={"type": "json_schema", "schema": TRIAGE_SCHEMA} ) ``` ### A2UI emission ```json {"version":"v0.9","updateComponents":{"surfaceId":"...","components":[ {"id":"urgency-badge","component":"Badge","color":"breaking","text":"BREAKING","pulse":true} ]}} {"version":"v0.9","updateDataModel":{"surfaceId":"...","path":"/triage","value":{...}}} ``` ืื urgency = "breaking" โ†’ ื”-frontend ืžื•ืกื™ืฃ **animation pulse ืื“ื•ื** ื‘-stepper. ื”-LLM ืœื ื™ื•ื“ืข ืขืœ pulse โ€” ื”-CSS ื™ื•ื“ืข (data-attribute `[data-urgency="breaking"]`). ### Cost - 1 Gemini 2.5 Flash call ยท ~500 input + ~200 output tokens - **Cost: ~$0.0003 per request** --- ## ๐Ÿ—‚๏ธ J04 ยท Desk Router ### ืชืคืงื™ื“ ืžืฆื ืืช ื”-**desk** ื”ื ื›ื•ืŸ ืžืชื•ืš 11: 1. **breaking** โ€” ื—ื“ืฉื•ืช ืžื™ื™ื“ื™ื•ืช 2. **security** โ€” ืฆื‘ื, ื‘ื™ื˜ื—ื•ืŸ, ื˜ืจื•ืจ 3. **politics** โ€” ื›ื ืกืช, ืžืžืฉืœื”, ืžืคืœื’ื•ืช 4. **economy** โ€” ื‘ื•ืจืกื”, ืžืืงืจื•, ืžื˜ื‘ืขื•ืช 5. **sports** โ€” ื›ื“ื•ืจื’ืœ, NBA, ืื•ืœื™ืžืคื™ืื“ื” 6. **tech** โ€” ื˜ื›ื ื•ืœื•ื’ื™ื”, ืกื™ื™ื‘ืจ, AI 7. **entertainment** โ€” ืงื•ืœื ื•ืข, ืžื•ื–ื™ืงื”, ืชืจื‘ื•ืช 8. **health** โ€” ืจืคื•ืื”, ื‘ืจื™ืื•ืช ื”ืฆื™ื‘ื•ืจ 9. **social** โ€” ื—ื‘ืจื”, ืจื•ื•ื—ื”, ื—ื™ื ื•ืš 10. **gallery** โ€” ื’ืœืจื™ื•ืช ืืžื ื•ืช, ืื•ืคื ื” 11. **archive** โ€” ื”ื™ืกื˜ื•ืจื™ื”, ืืจื›ื™ื•ืŸ ื”-desk ืงื•ื‘ืข: - **Color** ืฉืœ ื”-flow ื‘-UI (`var(--desk-{name})`) - **Writers pool** (J13-J16) โ€” ืื™ื–ื” writer agent ื™ื•ืงืฆื” - **Vector index** (J10) โ€” ืื™ื–ื” pgvector index ืœื—ืคืฉ - **Distribution channels** (J39-J44) โ€” TBD ### Implementation **Two-stage**: #### Stage 1: Classifier model (fast, 100ms) fine-tuned BERT classifier ืขืœ 50K ื›ืชื‘ื•ืช ืžืขื‘ืจ. ืžื—ื–ื™ืจ `top_3 desks + scores`. #### Stage 2: LLM tie-breaker (LLM, only if top_2 close) ืื `top_1.score - top_2.score < 0.15` โ†’ ืงืจื™ืื” ืœ-Gemini ืขื 2 ื”desks ื”ืžื•ื‘ื™ืœื™ื + ื”ื˜ืงืกื˜ โ†’ final decision. ### A2UI emission ```json {"version":"v0.9","updateComponents":{"surfaceId":"...","components":[ {"id":"desk-badge","component":"Label","variant":"filled","color":"security","text":"ื‘ื˜ื—ื•ืŸ"} ]}} {"version":"v0.9","updateDataModel":{"surfaceId":"...","path":"/classify/desk","value":{ "primary":"security", "scores":{"security":0.87,"politics":0.09,"breaking":0.03}, "tie_breaker_used":false }}} ``` ### Cost - BERT inference: ~$0.00001 (self-hosted on EX63 GPU) - LLM tie-breaker (15% of cases): ~$0.0002 - **Average per request: ~$0.00004** --- ## ๐Ÿ“Š Reception Quartet ยท Total | Station | Latency | Cost | Type | |---|---|---|---| | J01 Reception | <50ms | $0.000001 | Deterministic | | J02 Language | <200ms | $0.000005 | Heuristic + LLM fallback | | J03 Triage | <500ms | $0.0003 | LLM (Gemini Flash) | | J04 Desk Router | <800ms | $0.00004 | BERT + LLM tie-breaker | | **Total** | **<1.5s** | **~$0.00035** | | ื‘-100K ื”ื•ื“ืขื•ืช ื‘ื™ื•ื (ื’ื•ื“ืœ newsroom ืืจืฆื™): - **Latency p95**: 1.2 sec from paste to "ready for 5W" - **Daily cost (Reception only)**: $35 - **Monthly**: $1,050 --- ## ๐ŸŽฏ ื”ื”ื›ืจืขื•ืช ื”ืงืจื™ื˜ื™ื•ืช ### 1. ืœืžื” 4 stations ื•ืœื station ืื—ื“ ื’ื“ื•ืœ? 3 ืกื™ื‘ื•ืช: - **Observability**: ื›ืœ station ื™ื›ื•ืœ ืœื”ื™ื›ืฉืœ ืขืฆืžืื™ืช. ื ืจืื” ื‘-Lobby dashboard ืื™ื–ื” station ื ืชืงืข. - **Caching**: J04 (Desk Router) cacheable per-text-hash. ืื ืื•ืชื• ื˜ืงืกื˜ ื ืฉืœื— ืฉื•ื‘ โ†’ 0 latency, 0 cost. - **Replacement**: BBC ืจื•ืฆื™ื algorithm ืื—ืจ ืœ-Triage? ืžื—ืœื™ืคื™ื ืจืง J03, ื”ืฉืืจ ื ืฉืืจื™ื. ### 2. ืœืžื” J02 ืœืคื ื™ J03 ื•ืœื ื”ืคื•ืš? ื”triage prompt ืชืœื•ื™ ืฉืคื”. ืื ื”ืžืกืžืš ื‘ืขืจื‘ื™ืช, ื”prompt ืขื‘ืจื™ืช ืœื ื™ืขื‘ื•ื“ ื˜ื•ื‘. **Language first**. ### 3. ืœืžื” Desk Router ื‘-station ื ืคืจื“ ื•ืœื ื—ืœืง ืž-Triage? Triage = "ื›ืžื” ื“ื—ื•ืฃ". Desk Router = "ื‘ืื™ื–ื” ืชื—ื•ื". ืฉื ื™ concerns ืฉื•ื ื™ื. Triage ื™ื›ื•ืœ ืœื”ื™ื•ืช "breaking" + Desk = "sports" (ืืกื•ืŸ ื‘ื›ื“ื•ืจื’ืœ). ืฉื ื™ื”ื ืขืฆืžืื™ื™ื. ### 4. Embargo handling ืื J03 ืžืฆื embargo โ†’ ื”-flow ืžืžืชื™ืŸ ื‘ืžืฆื‘ `pending_embargo`: - ื›ืœ station ืื—ืจ ืขื•ื‘ื“ ืขืœื™ื• (5W extraction, context search, writing) โ€” ื”ื›ืœ ืžื•ื›ืŸ - ืื‘ืœ **Distribution (J39-J44) ื—ืกื•ื** ืขื“ `embargo.until` - ื‘-`embargo.until` โ†’ cron triggers `unembargo` event โ†’ distribution starts ื–ื” ื™ื™ื—ื•ื“ ืฉืœ newsroom workflow โ€” ืœื ืงื™ื™ื ื‘-A2UI base catalog. --- ## ๐Ÿ”Œ A2UI Catalog Entries ืœืงื•ืžื•ืช J01-J04 ```json { "components": { "ReceptionStatus": { "type": "object", "description": "J01 status indicator in Lobby", "properties": { "state": { "type": "string", "enum": ["idle", "received", "validating", "valid", "rejected"] }, "charCount": { "$ref": "#/$defs/DynamicNumber" }, "errorMsg": { "$ref": "#/$defs/DynamicString" } } }, "LanguageBadge": { "type": "object", "description": "J02 detected language indicator", "properties": { "lang": { "type": "string" }, "direction": { "type": "string", "enum": ["ltr", "rtl"] }, "confidence": { "$ref": "#/$defs/DynamicNumber" } } }, "UrgencyIndicator": { "type": "object", "description": "J03 triage urgency display", "properties": { "level": { "type": "string", "enum": ["breaking", "urgent", "standard", "archive"] }, "pulseAnimation": { "type": "boolean", "default": false }, "reasons": { "type": "array", "items": { "type": "string" } } } }, "DeskBadge": { "type": "object", "description": "J04 desk classification with color", "properties": { "desk": { "type": "string", "enum": ["breaking","security","politics","economy","sports","tech","entertainment","health","social","gallery","archive"] }, "confidence": { "$ref": "#/$defs/DynamicNumber" }, "tieBreakerUsed": { "type": "boolean" } } } } } ``` --- ## โœ… Closure - [x] J01 Reception spec'd (deterministic, <50ms) - [x] J02 Language Detector (heuristic + LLM fallback) - [x] J03 Triage (Gemini structured output) - [x] J04 Desk Router (BERT + LLM tie-breaker) - [x] A2UI catalog entries defined - [x] Cost + latency budget calculated - [x] Embargo workflow documented โœ… **Round 11 closed. 4 floors of 44+10 documented.** --- ## ๐Ÿ›ฃ๏ธ Next: Round 12 โ€” J05-J09 (5W Parallel Agents)