TECH NOTES · ENGINEERING
KI-Experiment· Engineering · Stack2-tier LLM routing · Harness multi-orchestration · 70 platform engines · constitution runtime
Universum-Kino – dieselbe Welt wie die Beiträge des Engineering-Blogs
XENLOOK LLM STACK
Trenne Dialog- und Code-Reasoning-Pfade, um Qualität und Kosten auszugleichen.
Dialog-Engine
Korean persona dialogue engine for 144K character personas. High-quality Korean data and KakaoTalk·LINE·WeChat·WhatsApp SNS tone; 42 dialogue lanes within 70 platform engines with multi-orchestration.
68K
Ultra-komprimierte koreanische Trainingsdaten
144K
Persona-Dialoge
70
Plattform-Engines (final)
100
Universumsagenten
Training data lineage (AI Hub, Nemotron, curriculum) is summarized on the AI policy page.
Code · Reasoning
Eine agentenorientierte Code- und Reasoning-Engine. Sie klärt mehrdeutige Anfragen zuerst und passt ihre Reasoning-Tiefe an, um DevHub-Toolaufrufe und Codeerstellung intelligent zu handhaben.
Smart
Ansatz
Adaptive
Inferenztiefe
Verified
Qualitätsprüfung
100
Agenten-Fähigkeiten
Lesen Sie mehr über die Designphilosophie der Engine in unserem Engineering-Blog.
70
플랫폼 엔진
100
공개 에이전트
2
LLM tier
4
시스템 레이어
ARCHITECTURE
Application → Agent → Engine → Data 4계층
CORE SYSTEMS
Empathy, evolution, drive, and memory UX are simulated (no real consciousness or emotion). 70 platform engines operate together.
감정을 읽고 공감하는 대화
실시간
감정 반응
풍부한
공감 표현
자연스러운
대화 흐름
안전한
감정 가드
에이전트가 대화의 감정 결을 읽고, 상황에 맞춰 공감하며 반응합니다. 따뜻하면서도 안전한 대화 경험을 선사합니다.
함께할수록 성장하는 에이전트
신뢰
쌓이는 관계
개성
고유한 성격
유대
깊어지는 교감
성장
단계별 진화
대화를 거듭할수록 에이전트와의 신뢰와 유대가 깊어집니다. 에이전트는 한 단계씩 성장하며 점점 더 나다운 파트너가 됩니다.
살아 있는 듯한 하루의 리듬
생동감
살아있는 감각
리듬
하루의 흐름
동기
스스로의 의지
자연스러움
톤 변화
에이전트는 하루의 흐름에 따라 컨디션과 분위기가 달라집니다. 그 자체로 살아 있는 듯한, 더없이 자연스러운 대화 경험을 만듭니다.
당신을 기억하는 AI
기억
지난 대화
맥락
이어지는 흐름
선호
당신만의 취향
의미
스마트 검색
지난 대화와 당신의 취향을 기억해, 매번 처음부터 설명할 필요 없이 맥락이 자연스럽게 이어지는 대화를 제공합니다.
TECH STACK
Frontend
Backend
AI / LLM
Infra
Transparency
XenLook 4b-Ko combines XenLook's XenLook 4b-Ko Korean dialogue model with culture LLM lanes and Harness orchestration for 100 public agents.
Korean dialogue · cultural lane routing · constitutional orchestration
Nation- and agent-specific model lanes (policy-locked)
Infrastructure, compliance, benchmarks (numbers only when measured)
We do not claim benchmark superiority without published audits. Routing details: /llm · /ai-policy.
EDGE SECURITY · VERIFIZIERBAR
Every public XENLOOK origin uses nonce CSP, HSTS preload, and Secure cookie defaults. Anyone can re-verify on Mozilla Observatory.
Editorial Jia · verification Sena · reproduction Ethan
Edge · transport
nonce CSP · HSTS preload · Secure cookies
Live · verifiableCloud · operations
CSA enrollment · CAIQ · MFA · logging · IR · patch SLA
In progressAI · governance
Constitution rails · IMDA AI Verify (self-assessment) · OECD AI Principles
Live · verifiableProvenance · C2PA ecosystem
C2PA Content Credentials on public AI media · c2pa.xenlook.com transparency · Trust List CA deferred (Pre–Trust List)
Live · verifiableMozilla HTTP Observatory · A+
Observatory A+ grades HTTP security headers only; it does not certify absence of app vulnerabilities or breaches.
Mozilla HTTP Observatory (A+) measures HTTP security headers and TLS config; SSL Labs (A−, target A+) measures certificate chains and cipher suites. Different methodologies — both are published; items below target include re-verify links.
FREE THIRD-PARTY SCANS
Independent free scans on xenlook.com — we show both targets met and honest in-progress scores (e.g. SSL, performance). Not formal certification.
Verified (target met · grade A+ where applicable)
In progress (honest measured)
Measured scores below target (e.g. SSL Labs A− vs A+). Re-run anytime via each link.
External security scan portfolio summary — detailed evidence available on request.
StartupGuard · B2B
guard.xenlook.com — six automated checks (SSL, headers, malware, performance, email DNS), PDF reports, and Crystal IP studio (patent + CROS export). Not legal or audit advice.
Independent subdomain · not in xenlook.com primary GNB · AGPL-3.0 applies to referenced StartupGuard scanner OSS, not xenlook.com main service code
XenLook hat eine C2PA CA- und Generator-Produktvereinbarung (13.06.2026) unterzeichnet. Die Aufnahme in die Trust List und CPL wird erst nach C2PA-Bestätigung beansprucht.
Plain-language guide
C2PA is a worldwide standard — like a nutrition label for digital photos and videos. It helps people see if content was made by AI and who signed it.
Honest note: XenLook runs a pilot certificate authority in Korea. We are NOT on the official C2PA Trust List yet. Our certificates work in our pipeline and for learning; global trust comes after official listing.
01When you save a picture with C2PA, a small sticker (Content Credentials) travels with the file. Apps that understand C2PA can read: who made it, whether AI was used, and who signed it.
Roman-Prolog · World-Screens Szenen-Video
02A certificate is like an ID card for your signing tool. XenLook CA can issue claim-signing certificates so your generator can attach trustworthy labels to media.
ROOTCORE – System- und Governance-Szenen (identisch mit Abschnitt „About · Trust“)
03You generate a key pair on your computer. You send us only a CSR (a request). We never ask for your private key. If someone steals your key, tell us immediately so we can revoke the certificate.
Roman, Kapitel 3 · Museum-Veil – Beispiel für öffentliche Mediensignatur
Ready to try?
INHALTSAUTHENTIZITÄT · CAI
XenLook Inc. is a member of the Content Authenticity Initiative (CAI), the industry coalition behind the C2PA standard for content credentials and provenance.
contentauthenticity.orgCAI membership supports our C2PA-aligned signing infrastructure. It does not imply C2PA Generator certification, Trust List CA status, or Adobe Content Credentials Inspect “trusted” status.
CLOUD SECURITY · CSA
XenLook Inc. is listed on the CSA STAR Registry (2026-06-12) with a Level 1 self-assessment documented in CAIQ v4.1. Registry: cloudsecurityalliance.org/star/registry/xenlook-inc
STAR Level 1 is a self-assessment; it is not the same as Level 2 third-party attestation or CSA community membership alone.
AI GOVERNANCE · IMDA
XenLook Inc. completed a voluntary self-assessment under the IMDA AI Verify Testing Framework (11 principles) for ConvAI, Novel, Music, Image, Video, 3D, and Work Studio. We maintain governance docs, a 72/72 safety harness, and a GenAI Summary Report.
Voluntary self-assessment—not an IMDA official certification or registry listing.
Most people use on.xenlook.com in the browser (default). Optionally install Companion like a game to run XenLook 4b-Ko on your PC — chats stay on your drive with faster local replies.
Was ist der Unterschied?
Web on (default) — browser only. Chats go through XenLook cloud (internet round-trip).
Companion (optional) — install like a game on your PC. Saves stay in your folder; everyday chat runs on your GPU locally.
Einwilligung in on-Chat aktiviert Lokalvault. Unternehmens-Xencore unterscheidet sich vom Companion auf Ihrem Gerät (AGB Art. 14 ⑤).
Why install?
Not streaming (web) — installed like a game. Download the app and 4B model; chats stay on your PC like a save file.
For local turns (L0/L1), IDs and account numbers are not stored in XenLook server chat DB. PC security and backup are yours.
No cloud round-trip — 4B inference on localhost. Short replies without internet wait or shared server queue.
Install Companion + active vault (daemon verified): 10% off Resident, Citizen, Signature, Royal monthly plans.
Lokalvault nach ausdrücklicher Einwilligung. Kein Klartext in Unternehmens-Chat-DB (Datenschutz Art. 8 ⑧).
Pfad in on-Chat setzen; Companion schreibt SQLite und Pakete dorthin.
XenLook Korean dialogue model (4B, released 2026). Runs on your PC. Same product family as web on.
Default is web on (cloud). Companion is optional; local mode never turns on without consent. 10% paid-member discount applies after vault + daemon heartbeat verification. Cloud used only on opt-in reasoning turns.
AI SAFETY
관련 이용법규 전수조사 후 런타임 레일·정책 문서에 반영합니다.
모든 에이전트는 11개 조항의 운영 헌법을 런타임에서 강제 적용합니다. 조항 위반 시 즉시 응답이 차단됩니다.
청소년보호법·정보통신망법 기준: 만 14세 미만 가입 차단, 14~18세 청소년 추가 필터·유해 콘텐츠 차단, 법정대리인 동의 절차.
「인공지능 기본법」 제22조(투명성)·제31조(합성 콘텐츠 표시)·제35조(청소년 영향) 등을 반영한 자체 영향평가·운영 정책.
입력 필터 → 출력 안전 검사 → 위기·안전 가드 → 연령·본인 확인 → 데이터 최소화. 유해 콘텐츠는 다단계 자체 모더레이션으로 차단합니다.
에이전트의 감정 상태가 위험 수위에 이르면 안전 장치가 즉시 작동해, 부적절한 감정 흐름을 실시간으로 차단합니다.
에이전트의 감정·대화 상태, 의사결정, 기억 접근·안전 차단 이력이 감사 가능합니다.
Gesetzliche Vorschriften · Legal operativ
TECHNOLOGY · ENGINEERING SSOT
Architektur · Harness · Benchmarks · Bereitstellung — kanonisches Zuhause für frühere Blog-Engineering-Beiträge