ArgMed — klinik mulohaza dvijoki

• Bu nima: US healthtech platformdan «yurak-qon tomir sohasi (TA1) uchun klinik mulohaza dvijoki». Bemorning klinik kontekstini oladi va ko'p agentli debat orqali gipotezalarni tushuntirish sifatining epistemik bahosi bilan yaratadi/tanqid qiladi/sintez qiladi. Klinik gaydlaynlar bo'yicha RAG, ko'p ijaralilik, audit. Deutsch (TA1) / Popper (TA2) / Hermes / PHI-servis ekotizimining bir qismi.
• Turi / holati / roli: api/dvijok (Bun monorepozitoriysi) / faol / lider — foydalanuvchi Davron Yuldashev <yul.davron.93@gmail.com> = 292 dan 170 kommit (123 «Davron Yuldashev» + 47 «Dave93») (~58%) va dvijok yadrosi paketining bosh muallifi (core'ga 28 tahrir, Anton Kim'da 18, Harsh'da 17, aniashev'da 11). Jamoa: Anton Kim, Harsh Manwani, Anna Shevtsova/aniashev.
• Faollik oynasi: 2026-01-26 → 2026-03-11 (~1,5 oy jadal ish), 292 kommit.

• Tillar: TypeScript (Bun runtime).
• **Monorepozitoriy (Turborepo + Bun vorkspeyslari @deutsch/*): `apps/`: api (Elysia), queue. `packages/`: core (mulohaza dvijoki), cartridges/cvd** (kardiologiya domen kartriji), db (Drizzle + TimescaleDB + pgvectorscale), client (TS SDK), adapters (popper TA2, phi), config-*.
• AI: Vercel AI SDK 6 ko'p-provayderli reyestr va HIPAA-aware failover bilan: Vercel Gateway → Azure OpenAI (BAA) → AWS Bedrock (BAA) → Anthropic → OpenAI; BYOK. Model aliaslari (reasoning-primary=Claude Sonnet 4.5, reasoning-fast/response=Haiku 4.5, embeddings text-embedding-3-large 3072d). Reyt-limitlar bilan model-registry.json, switch-provider.sh/model-query.sh, Cerebras/Vertex uchun presetlar.
• Ma'lumotlar: PostgreSQL 17 + TimescaleDB (gipertajadvallar: audit_events 6 yil saqlash, session_activity 90 kun, kompressiya/saqlash siyosatlari) + pgvectorscale/pgvector (RAG: guideline_embeddings, interaction_embeddings). Ko'p ijaralilik (tenants, sessions).
• Infra/deploy: Docker (apps/api/Dockerfile), KSA deploy (docker-compose.ksa.yml — Saudiya Arabistoni), GitHub Actions CI (lint/typecheck/test/build/docker), Biome.

Foydalanuvchi — yetakchi dasturchi; ham platformaga, ham dvijokning muhim qismiga egalik qiladi.

• ArgMed dvijok yadrosi (packages/core/src/engine/) — bosh muallif: debate-orchestrator, proposal-generator, claim-classifier, contradiction-detector, survivor-selector, confidence-calculator, htv-scorer, bold-rating, counter-hypothesis, idk-trigger, mode-enforcer, diversity-analyzer, snapshot-validator, session-manager, context-builder, output-validator.
• AI qatlami (packages/core/src/ai): provayderlar, embeddinglar, klient, ArgMed Zod-sxemalari (Generator/Verifier/Reasoner Output, HTVScore, ClaimType).
• CVD kartriji (kardiologiya domen bilimi) — bosh muallif.
• API/BD/navbatlar: Elysia API, Drizzle + TimescaleDB sxemasi, navbatlar paketi.
• Hajm: 170/292 kommit (~58%), shu jumladan dvijokning nostrukturaviy algoritmlari.

Kod bilan tasdiqlangan (packages/core/src/engine/*).

• ArgMed: uch agentli debat (debate-orchestrator.ts) → konfiguratsiya bilan Generator→Verifikator→Rezoner pipeline'i: htvThreshold (gipoteza qabul qilinishining minimal bali), maxRounds, retry bilan da'vo turlarini qamrab olish tekshiruvi (validateClaimTypeCoverage, retryAttempted/retrySucceeded), framework-agnostic metrik kolbeklari (PhaseMetricsCallback — Prometheus/OTel) va SSE-progress (PhaseStartCallback). Yetuk LLM-agent orkestratsiyasi.
• HTV-skoring (Hard-To-Vary) (htv-scorer.ts) → Devid Doychning yaxshi tushuntirish mezonini amalga oshirish: Interdependence / Specificity / Parsimony / Falsifiability o'qlari bo'yicha o'lchov; chegaralar refutation 0.3 / idk 0.4 / good 0.7 / excellent 0.85. Chegaradan past gipoteza → rad etiladi yoki IDK protokolini ishga tushiradi (gallyutsinatsiya o'rniga «bilmayman»). Bu kamdan-kam uchraydigan, ongli epistemik muhandislik (prodakshnda Popper/Doych).
• Falsifikatsiya orqali anti-gallyutsinatsiya → contradiction-detector, counter-hypothesis, survivor-selector, idk-trigger — tizim ishonchli yolg'on o'rniga yomon asoslangan gipotezalarni rad etadi. Klinik kontekst uchun juda muhim.
• HIPAA-aware ko'p-provayderlilik → BAA-provayderlar (Azure/Bedrock) bilan failover zanjiri, BYOK; purpose bo'yicha marshrutlash bilan deutschGenerateText/Object/Embed abstraksiyasi.
• Bitta Postgres'da time-series + vektor → TimescaleDB gipertajadvallari (saqlash/kompressiya siyosatlari) + RAG uchun pgvectorscale; saqlashlar zоoparki o'rniga bitta baza.

• Topilgan sessiyalar: loyiha uchun lokal Claude Code sessiyalar katalogi mavjud, ammo unda 0 `.jsonl`-transkript bor (tozalangan/saqlanmagan).
• Bilvosita: batafsil CLAUDE.md (11 KB), CONTRIBUTING.md, zich hujjatlar (kod ishora qiladigan docs/01-architecture/02-argmed-framework.md) — AI yordamida, muhandislik intizomi bilan ishlab chiqish.
• AI-vorkflow patternlari: transkriptlar yo'q; ammo repozitoriy «AI-driven, lekin qat'iy spetsifikatsiyalangan» ishlab chiqishning namunaviy misoli (Zod-sxemalar, har bir dvijok moduli uchun testlar: htv-scorer, debate, claim-classifier, survivor-selector va h.k. uchun *.test.ts).

• Foydalanuvchi: 170 kommit (~58%), dvijok yadrosining yetakchi muallifi.
• Dvijok: 15+ mulohaza moduli + uch agentli debat + 4 o'q bo'yicha HTV-skoring + IDK protokoli.
• BD: 6 jadval, 2 gipertajadval (6 yillik audit), 2 RAG vektor jadvali, ko'p ijaralilik.
• AI: 5 bosqichli HIPAA-failover, 6 model aliasi, reyt-limitli reyestr, Cerebras/Vertex/Azure/Bedrock presetlari.
• Dvijok test qamrovi (har bir modul uchun bun:test).
• KSA mintaqasida deploy.