هيكلية OmniDiag: منصة تشخيص سريري متعددة الأمراض

Modular Monolith بعملية FastAPI واحدة وواجهة React. ١٣ مخططاً تغطي المعمارية، تدفق الطلب، النماذج، الإنسان في الحلقة، NLP والـ LLM، الأمن، المراقبة، قاعدة البيانات، الواجهة، النشر، والتعلّم الاتحادي.

BackendFastAPI, SQLAlchemy async, Alembic
FrontendReact, Vite, Tailwind, PWA
Diseasesheart_disease, diabetes_nhanes (BRFSS diabetes retired: 410)
StateSQLite dev / PostgreSQL, Redis, MLflow
OpsDocker, HF Space, Vercel, k8s, Helm, Prometheus, Grafana
Testsabout 60 pytest files
الحالة الصادقة: التعلّم الاتحادي ليس جزءاً من المنتج المشغَّل (القسم ١٣)، وترقية النموذج المرشَّح بعد إعادة التدريب قرار بشري لا يوجد كود يفعله تلقائياً (القسم ٦).

١. الصورة الكبرى: Modular Monolith

عملية FastAPI واحدة (deployable واحد) مقسّمة داخلياً إلى وحدات (modules) بحدود واضحة؛ الأمراض تُضاف بملف YAML دون لمس باقي الكود.

flowchart TB
  subgraph CL["Clients"]
    SPA["React SPA - Vite, Tailwind, PWA<br/>(Vercel)"]
    EMR["EMR / API clients<br/>JWT cookie or X-API-Key"]
  end
  subgraph MW["Cross-cutting middleware (order of wrapping)"]
    direction LR
    CORS["CORS"] --> SEC["SecurityHeaders<br/>HTTPS enforce"] --> AUD["AuditMiddleware<br/>writes audit_logs"] --> MET["MetricsMiddleware<br/>Prometheus"] --> RL["slowapi rate limit<br/>per-endpoint"]
  end
  subgraph APP["FastAPI process = the monolith (backend/main.py)"]
    subgraph MODS["Feature modules (APIRouter each)"]
      AUTH["auth<br/>/auth"]
      ADM["admin<br/>/admin"]
      PAT["patients<br/>/api/v4/patients"]
      CLIN["clinical diagnosis<br/>/api/v4/{disease}/..."]
      AL["active_learning (HITL)<br/>/api/v4/review"]
      MON["monitoring<br/>/api/v4/admin/drift, mlflow"]
      NLP["nlp<br/>notes_parser"]
      LLM["llm<br/>report_generator"]
      FED["federated<br/>SUPERSEDED prototype"]
    end
    subgraph CORE["Shared core services"]
      ROUTER["OmniDiagRouter<br/>scans configs/*.yaml"]
      PS["probability_scale +<br/>prevalence_correction"]
      SHAP["shap_service"]
      CF["counterfactual_generator +<br/>diabetes_what_if_levers"]
      CACHE["cache.py<br/>Redis or in-memory"]
    end
    subgraph REG["Model-family registry (backend/model_backends)"]
      B1["glm_ivap_conformal<br/>HEART"]
      B3["ebm_platt_conformal<br/>DIABETES NHANES"]
      B4["sklearn_pipeline /<br/>sklearn_generic"]
    end
  end
  subgraph DATA["State and artifacts"]
    DB[("SQL DB<br/>SQLite dev / PostgreSQL prod<br/>SQLAlchemy async + Alembic")]
    RED[("Redis cache")]
    ART[("models/ bundles<br/>built at image build")]
    MLF[("MLflow<br/>runs and registry")]
    CFG["configs/*.yaml<br/>single source of truth per disease"]
  end
  subgraph OBS["Observability"]
    PROM["Prometheus"] --> GRAF["Grafana"]
  end
  subgraph EXT["External"]
    DS["DeepSeek API<br/>(OpenAI-compatible)"]
    BERT["BioBERT NER<br/>HuggingFace, optional"]
  end
  CL --> MW --> MODS
  CLIN --> ROUTER --> REG
  CLIN --> PS
  CLIN --> SHAP
  CLIN --> CF
  CLIN --> CACHE
  CACHE -.-> RED
  CFG --> ROUTER
  REG --> ART
  AUTH --> DB
  PAT --> DB
  ADM --> DB
  AL --> DB
  AUD -.-> DB
  MON --> MLF
  AL --> MLF
  MET --> PROM
  LLM -.-> DS
  NLP -.-> BERT

٢. خريطة الوحدات (Package Map)

كل مجلد = وحدة بمسؤولية واحدة. الوحدات تتحدث عبر استدعاءات داخلية (in-process) لا عبر شبكة، وهذا ما يجعله Monolith، والحدود النظيفة بينها هي ما يجعله Modular.

flowchart LR
  ROOT(["OmniDiag repo"])
  ROOT --> BE["backend/"]
  ROOT --> FE["frontend/src"]
  ROOT --> CF["configs/ + features/"]
  ROOT --> SC["scripts/ + models/ + experiment_files/"]
  ROOT --> OPS["Dockerfile, docker-compose, k8s, helm, deploy, .github"]
  ROOT --> TST["tests/ (about 60 files)"]
  ROOT --> DOC["docs/, evaluation_evidence/, alembic/"]
  BE --> m1["main.py - app wiring, lifespan,<br/>predict/explain/batch/report/parse-notes"]
  BE --> m2["router.py - disease dispatcher"]
  BE --> m3["model_backends/ - 4 families,<br/>auto-registered by file"]
  BE --> m4["heart_glm/ - stack.py, candidate.py,<br/>reference_scores.json"]
  BE --> m5["auth/ - jwt, hashing, rbac, api_key"]
  BE --> m6["admin/ - users, audit, cache flush,<br/>api-keys, stats, retrain"]
  BE --> m7["patients/ - EMR: patients, visits,<br/>predictions, export, notes"]
  BE --> m8["active_learning/ - sampler, routes,<br/>retrain, nhanes candidate"]
  BE --> m9["monitoring/ - drift, drift_stats,<br/>metrics, mlflow_tracker"]
  BE --> m10["nlp/ + llm/"]
  BE --> m11["middleware/ - audit, security"]
  BE --> m12["db_models/ - 8 ORM tables"]
  BE --> m13["federated/ - superseded"]
  CF --> c1["heart_disease.yaml<br/>diabetes_nhanes.yaml"]
  CF --> c2["base_features.py<br/>(extension point; no live module uses it)"]
  SC --> s1["train_heart_glm.py<br/>build_drift_reference.py<br/>retrain.py, seed_db.py"]

٣. رحلة طلب التنبؤ (Predict Request)

ما يحدث من الضغط على Predict حتى ظهور النتيجة: المصادقة، الكاش، النموذج، التسجيل، وإضافة الحالات غير المؤكدة لطابور المراجعة البشرية.

sequenceDiagram
  autonumber
  participant U as Clinician (React)
  participant MW as Middleware
  participant API as POST /api/v4/{disease}/predict
  participant AU as Auth + RBAC
  participant C as Cache (Redis or memory)
  participant R as OmniDiagRouter
  participant B as ModelBackend (per family)
  participant DB as SQL DB
  participant AL as Active-learning sampler
  U->>MW: JSON features
  MW->>API: rate limit, security headers, start timer
  API->>AU: JWT cookie or API key
  AU-->>API: user, roles must be doctor, nurse or super_admin
  API->>API: validate against schema from YAML
  API->>C: get(hash of disease + features)
  alt cache hit
    C-->>API: stored result (scale checked)
  else cache miss
    API->>R: predict in threadpool
    R->>B: engineer features, model, calibrate
    B-->>R: probability, decision, bounds
    R-->>API: result
    API->>C: set(result, ttl)
  end
  API->>API: record Prometheus counters (hit or miss)
  API->>DB: INSERT predictions (scale and threshold stored)
  API->>AL: uncertain?
  alt conformal decision is uncertain, or entropy near threshold
    AL->>DB: INSERT review_queue (pending)
  end
  API-->>MW: PredictResponse + Cache-Hit header
  MW->>DB: INSERT audit_logs (user, endpoint, status, ms)
  MW-->>U: result, SHAP available via /explain

٤. الإضافة بالإعدادات: Config-driven Registry

إضافة مرض = ملف YAML (وملف features فقط إن احتاجته العائلة). الـ Router لا يعرف أسماء الأمراض ولا العائلات. كل عائلة نموذج تسجّل نفسها بمجرد وجود ملفها في model_backends.

flowchart LR
  subgraph FS["Filesystem"]
    Y1["configs/heart_disease.yaml<br/>family: glm_ivap_conformal"]
    Y3["configs/diabetes_nhanes.yaml<br/>family: ebm_platt_conformal"]
  end
  subgraph STARTUP["At startup"]
    SCAN["OmniDiagRouter._load_all_configs()<br/>one entry per YAML"]
    IMP["model_backends/__init__.py<br/>pkgutil imports every module"]
    REGI[("registry<br/>@register_backend('name')")]
    IMP --> REGI
  end
  Y1 --> SCAN
  Y3 --> SCAN
  SCAN -->|"family key"| GET["get_backend(family)"]
  REGI --> GET
  GET -->|"unknown family fails at startup,<br/>names the registered ones"| ERR["UnknownModelFamilyError"]
  GET --> L1["HeartGlmConformal"]
  GET --> L3["EbmPlattConformal"]
  GET -.-> L4["sklearn_pipeline / sklearn_generic"]
  L1 -.->|"lazy load on first request"| W1[("heart_l3_glm_stack.pkl")]
  L3 -.-> W3[("diabetes_nhanes_ebm.joblib")]
  subgraph IFACE["ModelBackend interface"]
    I1["predict_proba, feature_names, shap_values"]
    I2["predict, predict_batch, explain"]
    I3["BackendCapabilities:<br/>tree_shap, vectorized_batch,<br/>counterfactuals, explainer"]
  end
  L1 --- IFACE

٥. نموذج القلب: Spline-GLM + Venn-Abers + Conformal

القرار ليس احتمالاً مقابل عتبة واحدة، بل مجموعة conformal تعطي referral / no_referral / uncertain، مع فاصل [p_lower, p_upper]. 7 مدخلات فقط، مدرَّب على 920 مريضاً من 4 مستشفيات.

flowchart LR
  IN["7 inputs (L3)<br/>Age, Sex, ChestPainType, RestingBP,<br/>Cholesterol, FastingBS, RestingECG"]
  PRE["Frozen preprocessing<br/>IterativeImputer, scaler,<br/>SplineTransformer, one-hot"]
  GLM["Spline-GLM<br/>logistic, ridge"]
  S["raw score<br/>(log-odds, exact additive SHAP)"]
  IVAP["IVAP<br/>inductive Venn-Abers"]
  P["calibrated p<br/>+ interval [p_lower, p_upper]"]
  MC["Mondrian conformal<br/>groups: Sex x class<br/>alpha = 0.10"]
  DEC{"decision"}
  R1["referral"]
  R2["no_referral"]
  R3["uncertain"]
  Q["review_queue<br/>(human in the loop)"]
  IN --> PRE --> GLM --> S
  S --> IVAP --> P
  S --> MC --> DEC
  DEC --> R1
  DEC --> R2
  DEC --> R3 --> Q
  subgraph LIM["Declared limits shown in config"]
    L1["Scope: patients already referred for catheterisation"]
    L2["No HIGH/MODERATE/LOW bands (risk_bands: null)"]
    L3["LOHO AUC 0.802, calibration does not transport between hospitals"]
  end
  DEC -.-> LIM

٦. الإنسان في الحلقة والتعلّم النشط (HITL + Active Learning)

النظام لا يتعلّم من نفسه تلقائياً: الحالات غير المؤكدة تذهب لطبيب، تسميته تُخزَّن، والإعادة التدريبية تنتج مرشَّحاً (candidate) يقرّر إنسان ترقيته.

flowchart TB
  P["Prediction"] --> Q1{"module type"}
  Q1 -->|"conformal (heart, NHANES)"| C1["decision == uncertain"]
  Q1 -->|"threshold-only module<br/>(none live; the retired BRFSS was one)"| C2["threshold-centred entropy >= 0.88<br/>(prior-shift map sends threshold to 0.5)"]
  C1 --> QUEUE
  C2 --> QUEUE
  QUEUE[("review_queue<br/>status = pending<br/>uncertainty_score, scale, threshold")]
  QUEUE --> UI["ReviewQueuePanel (doctor)<br/>GET /api/v4/review/queue"]
  UI --> ACT{"doctor action"}
  ACT -->|"annotate: label 0/1 + notes"| REV["status = reviewed"]
  ACT -->|"skip"| SKP["status = skipped"]
  REV --> RT["super_admin: POST /admin/retrain<br/>min-samples gate"]
  RT --> FAM{"model.family"}
  FAM -->|"heart: glm_ivap_conformal"| H1["build_bundle() on 920 UCI rows<br/>+ reviewed rows<br/>(same code as shipped build)"]
  H1 --> H2["models/heart_disease/candidates/RUN_ID"]
  H2 --> H3["compare to SHIPPED model:<br/>conformal coverage +<br/>decision transition matrix"]
  H3 --> H4["log to MLflow"]
  H4 --> H5{{"HUMAN decides promotion<br/>(no code path promotes)"}}
  FAM -->|"NHANES: ebm_platt_conformal"| N1["status: unsupported (not wired yet);<br/>diabetes_nhanes_candidate.py exists:<br/>candidate only, HbA1c provenance required"]
  FAM -->|"any other family"| D1["status: unsupported<br/>(nothing read or written)"]
  FAM -->|"retired disease (diabetes)"| D2["410 DISEASE_RETIRED"]
  SKP -.-> STATS["GET /review/stats"]
  REV -.-> STATS

٧. معالجة اللغة (NLP) والتقرير بالـ LLM

ملاحظة سريرية نصية تتحول إلى حقول جاهزة للنموذج؛ وبعد التنبؤ يُولَّد تقرير سريري بالـ LLM مع حارس يمنع المحتوى الممنوع وبديل قاعدي عند الفشل.

flowchart TB
  subgraph NLPX["NLP: POST /api/v4/parse-notes (20/min, no auth) or /patients/ID/notes"]
    N0["free-text note"] --> N1{"detect_script"}
    N1 -->|"Arabic"| N1a["language_support warning:<br/>patterns are English only"]
    N1 -->|"Latin"| N2["1. Regex baseline<br/>spelled numbers to digits, negation"]
    N2 --> N3["2. spaCy dependency parse<br/>numeric mentions, linking words,<br/>overrides regex (optional dep)"]
    N3 --> N4["3. BioBERT NER via HuggingFace<br/>lazy load, overrides on overlap"]
    N4 --> N5["map_to_disease_schema<br/>(heart only; NHANES has no map)"]
    N5 --> N6["pre-filled form<br/>missing fields omitted, user confirms"]
  end
  N6 --> PRED["/predict"]
  subgraph LLMX["LLM: POST /api/v4/generate-report (10/min)"]
    PRED --> L0["result: probability, decision, bounds, SHAP top 5"]
    L0 --> L1["band_for_report<br/>one band value used everywhere"]
    L1 --> L2{"DEEPSEEK_API_KEY set?"}
    L2 -->|"yes"| L3["DeepSeek chat via OpenAI SDK"]
    L3 --> L4{"forbidden_content(text)?"}
    L4 -->|"clean"| L5["source = llm"]
    L4 -->|"violations"| L6["rule-based template<br/>fallback_reason recorded"]
    L2 -->|"no"| L6
    L6 --> L7["source = rule_based"]
  end
  L5 --> OUT["ClinicalReportModal / PDFReport"]
  L7 --> OUT

٨. الأمن والصلاحيات والتدقيق

JWT في كوكي + Refresh، مفتاح API بديل للأنظمة الخارجية، صلاحيات RBAC، وسجل تدقيق لكل طلب.

flowchart TB
  subgraph AUTHF["Authentication"]
    REG["POST /auth/register"] --> HASH["password hashing"]
    LOG["POST /auth/login"] --> JWT["access JWT + refresh cookie"]
    REF["POST /auth/refresh"] --> JWT
    OUT["POST /auth/logout"]
    KEY["X-API-Key (hash stored)<br/>POST /admin/api-keys"]
  end
  subgraph AUTHZ["Authorisation (RBAC)"]
    DEP["get_current_active_user"]
    RR["require_role(...)"]
    ROLES["roles: doctor, nurse, super_admin<br/>CLINICAL_ROLES = doctor, nurse, super_admin<br/>ADMIN_ROLES = super_admin"]
    DEP --> RR --> ROLES
  end
  JWT --> DEP
  KEY --> DEP
  subgraph PROT["What each role reaches"]
    P1["predict, explain, patients, review: clinical roles"]
    P2["users, audit-logs, cache flush, retrain, drift run, mlflow: super_admin"]
  end
  RR --> P1
  RR --> P2
  subgraph DEF["Defence in depth"]
    D1["SecurityHeadersMiddleware + HTTPS enforcement"]
    D2["slowapi limits: LIMIT_CLINICAL, LIMIT_ADMIN, 10/min report, 20/min notes"]
    D3["AuditMiddleware: user, endpoint, method, status, IP, duration_ms"]
    D4["patient soft-delete (deleted_at)"]
  end
  D3 --> AL[("audit_logs")]

٩. المراقبة والـ Drift و MLflow

مراقبة تشغيلية (Prometheus/Grafana) ومراقبة انحراف المدخلات بإحصاءات KS و chi-square و PSI مقابل مرجع مجمَّد. الانحراف في المدخلات تنبيه للفحص وليس دليلاً على تدهور النموذج.

flowchart LR
  subgraph RUNTIME["Runtime metrics"]
    M1["MetricsMiddleware<br/>request duration, active requests"]
    M2["record_prediction<br/>predictions_total, confidence histogram"]
    M3["cache hits / misses"]
    M4["conformal decisions, batch rows"]
    M5["drift_share gauge"]
  end
  M1 --> ME["GET /metrics"]
  M2 --> ME
  M3 --> ME
  M4 --> ME
  M5 --> ME
  ME --> PR["Prometheus<br/>deploy/prometheus.yml"] --> GR["Grafana dashboards"]
  subgraph DRIFT["Input drift (Gate 8.7c)"]
    REFP[("drift_reference.json<br/>frozen profile<br/>built by build_drift_reference.py")]
    RECENT["recent predictions from DB"]
    PM["ProfileDriftMonitor<br/>KS, chi-square, PSI on scipy"]
    REFP --> PM
    RECENT --> PM
    PM --> ST["GET /drift/DISEASE/status"]
    PM --> RUN["POST /drift/DISEASE/run"]
    PM --> REP["GET /drift/DISEASE/report"]
    RUN --> M5
  end
  subgraph ML["MLflow"]
    TR["mlflow_tracker<br/>experiment OmniDiag"]
    TR --> RUNS["GET /mlflow/runs"]
    TR --> RG["POST /mlflow/register"]
  end
  RUN --> TR
  RET["retrain candidates"] --> TR

١٠. نموذج البيانات (Database)

ثمانية جداول؛ كل تنبؤ يُسجَّل مع مقياسه وعتبته، وطابور المراجعة مرتبط بالتنبؤ بعلاقة واحد لواحد.

erDiagram
  USERS ||--o{ USER_ROLES : has
  ROLES ||--o{ USER_ROLES : grants
  USERS ||--o{ PATIENTS : created_by
  PATIENTS ||--o{ PATIENT_VISITS : has
  PATIENTS ||--o{ PREDICTIONS : linked
  USERS ||--o{ PREDICTIONS : created_by
  PREDICTIONS ||--o| REVIEW_QUEUE : flagged
  USERS ||--o{ REVIEW_QUEUE : reviewer
  USERS ||--o{ AUDIT_LOGS : actor
  USERS {
    string id PK
    string email
    string hashed_password
    string api_key_hash
    bool is_active
  }
  ROLES {
    int id PK
    string name
  }
  PATIENTS {
    string id PK
    string mrn UK
    string full_name
    date date_of_birth
    datetime deleted_at
  }
  PATIENT_VISITS {
    string id PK
    string patient_id FK
  }
  PREDICTIONS {
    string id PK
    string disease
    float confidence
    string probability_scale
    string decision
    float probability_lower
    float probability_upper
  }
  REVIEW_QUEUE {
    string id PK
    string prediction_id UK
    float uncertainty_score
    float decision_threshold
    string status
    int label
    string notes
  }
  AUDIT_LOGS {
    string id PK
    string endpoint
    string method
    int status_code
    string ip_address
    int duration_ms
  }

١١. الواجهة الأمامية (React SPA)

الفورم يُبنى ديناميكياً من الـ schema القادم من الـ API (لا حقول مكتوبة يدوياً لكل مرض)، ولذلك تظهر الأمراض الجديدة بدون تعديل الواجهة.

flowchart TB
  MAIN["main.jsx"] --> PROV["Providers: Auth, Disease, Theme"]
  PROV --> APP["App.jsx - DiseaseSelector, theme toggle, view tabs"]
  APP --> V1["Engineering Mode"]
  APP --> V2["Clinical EMR Mode"]
  APP --> V3["Batch Prediction (CSV, max 500 rows)"]
  APP --> V4["Patient Comparison"]
  APP --> V5["Review Queue (clinical roles)"]
  APP --> V6["AdminDashboard"]
  subgraph SCH["Schema-driven form"]
    H1["useDiseaseSchema<br/>GET /api/v4/DISEASE/schema"]
    H2["schemaFieldParser + schemaToZod"]
    H3["DynamicClinicalForm + SchemaFieldFactory"]
    H4["ClinicalNotesInput (NLP prefill)"]
    H1 --> H2 --> H3
    H4 --> H3
  end
  V1 --> SCH
  V2 --> SCH
  subgraph RES["Result views"]
    R1["ThresholdBar + decision"]
    R2["ShapBarChart (useShapSnapshot)"]
    R3["WhatIfScenarioCard<br/>counterfactuals.js + /counterfactuals"]
    R4["ClinicalActionPlan"]
    R5["ClinicalReportModal + PDFReport"]
    R6["PatientRiskTimeline / PatientTimeline"]
  end
  SCH --> RES
  V5 --> RL["ReviewLabelBox<br/>annotate or skip"]
  APP --> PWA["PWA: InstallPrompt, OfflineBanner"]
  API["api.js - fetch wrapper, cookies"] --- SCH
  API --- RES
  API --- V5
  API --- V6

١٢. النشر والبنية التحتية (Deployment / CI)

نفس الصورة (image) تعمل محلياً وعلى Hugging Face Spaces وKubernetes. النماذج تُبنى أو تُحمَّل وقت بناء الصورة، فلا ملفات ثنائية في الريبو.

flowchart LR
  DEV["Developer"] --> GH["GitHub repo"]
  GH --> CI["CI: .github/workflows/ci.yml<br/>1. pip install, import backend<br/>2. start API check<br/>3. npm ci, npm run build"]
  GH --> DF["Dockerfile<br/>python 3.13-slim"]
  DF --> BUILD["Build time:<br/>train_heart_glm.py (CSV sha256 verified)<br/>download NHANES bundle from HF Hub<br/>(sha256-checked)"]
  BUILD --> IMG["Image: model baked in<br/>port 7860, instant start"]
  IMG --> HF["Hugging Face Space<br/>yahyoha/omnidiag"]
  IMG --> K8["Kubernetes<br/>k8s/*.yaml, helm chart,<br/>HPA, ingress, secrets"]
  IMG --> DC["docker-compose"]
  subgraph COMPOSE["docker-compose services"]
    PG[("postgres 15")]
    RD[("redis 7")]
    BE["backend"]
    MF["mlflow"]
    PRM["prometheus"]
    GRF["grafana"]
    RTN["retrain (profile: retrain)<br/>scripts/retrain.py"]
  end
  DC --> COMPOSE
  BE --> PG
  BE --> RD
  BE --> MF
  PRM --> BE
  GRF --> PRM
  RTN --> PG
  RTN --> MF
  FE["frontend build (vite dist)"] --> VC["Vercel<br/>SPA rewrite to index.html"]
  VC -->|"HTTPS API calls"| HF

١٣. التعلّم الاتحادي (Federated Learning): الحالة الفعلية

غير مُشغَّل في المنتج. النموذج الأولي (Flower + FedAvg + XGBoost) لا يستطيع التجميع، والبديل GLORE مُنفَّذ ومقاس في مستودع التجارب فقط.

flowchart LR
  subgraph OLD["backend/federated/ (SUPERSEDED, not used)"]
    O1["client.py<br/>pickle.dumps(XGBoost)"]
    O2["aggregator.py<br/>Flower FedAvg"]
    O3["add_dp_noise()<br/>never called, no epsilon"]
    O1 -->|"opaque bytes, cannot be averaged"| O2
    O2 -.-x O3
  end
  subgraph NEW["omnidiag_experiments (research only, not merged)"]
    G0["4 sites: Cleveland, Hungarian,<br/>Switzerland, Long Beach VA"]
    G1["each site computes locally<br/>gradient g and Hessian H"]
    G2["coordinator sums g, H<br/>arrow-block Newton step<br/>per-site intercept, shared slopes"]
    G3["7 iterations =<br/>centralised fit (error 2e-14)"]
    G0 --> G1 --> G2 --> G3
    G2 -->|"updated beta"| G1
  end
  NEW --> RES["Measured: LOSO AUC 0.779 vs 0.802 shipped (-0.023)<br/>no differential privacy, simulation on one machine"]