Deutsch Cookbook

• Bu nima: @deutsch/core dvigatelini namoyish qiluvchi va sinovdan o'tkazuvchi bajariladigan misollar bilan hamroh paket (bun run examples/NN-*.ts). Bu dvigatelning o'zi EMAS — bu uning API'si bo'yicha "cookbook": ichki komponentlardan (HTV skorlash, IDK, ishonchlilik, survivor selection) to to'liq servis yuzasigacha (sessiya API CRUD/TTL, SSE streaming, nazorat, audit, rate limiting, kalit rotatsiyasi).
• Turi / holati / roli: library (misollar/hujjatlar paketi) · faol · lead — 24 kommitdan 18 tasi foydalanuvchi tomonidan, aniashev co-author (6 ta).
• Faollik davri: intensiv sprint 2026-01-28 → 2026-02-02 (5 kun, har bir SAL-7xx/8xx tiketi bo'yicha kuniga ~bitta misol).

• Tillar: TypeScript (ESM, type: module).
• Freymvorklar/kutubxonalar: Bun (runtime/runner), @deutsch/core workspace bog'liqligi sifatida (workspace:*); @types/bun. tsconfig ota core'ga yo'naltirilgan path alias'lar bilan.
• Infra/deploy: yo'q — bular lokal ishga tushiriladigan skriptlar (bun run NN), o'rnatish talab qilinmaydi.
• Ma'lumotlar: o'ziga xos ma'lumotlari yo'q; dvigatelning tiplari/kontraktlari ustida ishlaydi (ArgMedHTVScore, Hypothesis, EngineContext, Cartridge, SnapshotRef va h.k.).
• E'tiborga molik tooling: dvigatelni integratsion sinovdan o'tkazish uchun mock agent fabrikalari.

Foydalanuvchining 18 ta kommiti — har biri aniq bir tiket uchun misol qo'shadi (misol kodidan xulosa qilingan):

• Dvigatel (epistemologiya): 05 HTV skorlash, 06 IDK protokoli, 07 ishonchlilik, 04 survivor selection, 08 to'liq debate (Generator→Verifier→Reasoner), 09 dvigatelni integratsion sinash (SAL-733).
• Rejimga bog'liqlik: rejimga sezgir IDK javoblari (SAL-718), klinik triaj metama'lumotlari (SAL-719), HTVScorer klassi (SAL-708).
• Sessiya API: create/get/delete (SAL-740..742), TTL majburlash (SAL-744), cleanup worker (SAL-745).
• Xabar almashish: send message (SAL-746), SSE streaming (SAL-747, 751..757), attachment'lar (SAL-748), snapshot refresh (SAL-750).
• Ops/observability: health/probe'lar/metrikalar (19), rate limiting, audit emitter (SAL-840..845), API kalit rotatsiyasi (SAL-876).
• Nazorat (human-in-the-loop): request builder (SAL-816..820), response parser (SAL-821..827), control command handler (SAL-828..834), fallback handler (SAL-835..839).
• Ko'lam: 5 kunda 24 kommit; PR oqimi (Merge PR #1 from US healthtech platform/SAL-821-827). aniashev — co-author.

Misol kodi bo'yicha (murakkab dvigatelni tushunishni namoyish etadi; mualliflik — foydalanuvchi):

• Popper bo'yicha HTV (Hard-To-Vary) skorlash (05-htv-scoring, 09-engine-testing): 4 o'lchov bo'yicha kompozit faraz skori (specificity 0.3 + falsifiability 0.3 + interdependence 0.2 + parsimony 0.2), chegaralar DEFAULT_HTV_THRESHOLDS (refutation 0.3 / idk 0.4 / good 0.7 / excellent 0.85). → Deutsch/Popper epistemologiyasini (yaxshi tushuntirish = o'zgartirish qiyin) klinik faraz baholashga tatbiq etish. Ahamiyatsiz bo'lmagan soha.
• IDK protokoli (mas'uliyatli AI) (06-idk-protocol): omon qolgan farazlar yo'q bo'lganda / past HTV / ziddiyat / yetarli bo'lmagan ma'lumotlarda dvigatel halol tarzda "bilmayman" deydi; rejimga sezgir ohang (wellness vs advocate_clinical) + klinik rejim uchun triaj metama'lumotlari (urgency/route_to/requires_supervision). → Noaniqlikni oshkor qilish va klinisistga eskalatsiya qilish uchun yetuk pattern.
• Mock agent fabrikalari orqali dvigatelni integratsion sinash (09): senariy presetlari bilan AgentFactory (happy_path / all_refuted / low_htv), mock Generator/Verifier + haqiqiy Reasoner, dalilga asoslangan maydonli data fabrikalari (createHTVScore, createHypothesis, createCritique) (ACC/AHA guideline_class, evidence_level). → Murakkab agentik tizimning sinovga yaroqliligiga puxta yondashuv.
• To'liq servis yuzasi (10–25 misollar): sessiya lifecycle + TTL/cleanup, SSE streaming, attachment'lar, observability endpoint'lari, rate limiting, audit, kalit rotatsiyasi, inson nazorati (request/response/control/fallback). → Faqat "modellar" emas, balki production darajasidagi AI servisni boshdan-oxir tushunishni namoyish etadi.

• Topilgan sessiyalar: bu katalog uchun 0 ta (normallashtirilgan kalit mavjud emas; to'liq yo'lni normallashtirish orqali tekshirildi). Ish, ehtimol, ota monorepo / asosiy deutsch repozitoriysidan bajarilgan (#19 masalaga qarang) — sessiyalarni o'sha yerda izlash kerak.
• AI bilan nima qilingan: bu katalogda ma'lumot yo'q; CLAUDE.md batafsil (220 qator) — ekotizimda AI yordamida ishlab chiqilganining belgisi.
• AI vorkflow patternlari: dvigatel uchun ma'lumotnoma sifatida batafsil CLAUDE.md; tiketga yo'naltirilgan kommitlar (SAL-xxx) — tuzilgan jarayon.

• Dvigatel + servisni boshdan-oxir qamrab oluvchi ~26 ta bajariladigan misol; tiket qamrovi SAL-708..876.
• Human-in-the-loop nazorati va auditi bilan production klinik AI tizimining (Healthcare platform/US healthtech platform) bir qismi.