Architecture

Scores architecture risk from a declared system inventory. HTTP service and CLI over one scoring core.

  PredictArchitectureRequest
  (companyName, systems[], historicalMetrics?)
            │
            ▼
  ┌──────────────────────────────────────────────┐
  │  CloudSealed.ML.Engine                         │
  │                                                │
  │  Scoring/RiskRules.cs   ── named weights &     │
  │        │                   thresholds, each    │
  │        │                   with a rationale    │
  │        ▼                                        │
  │  Scoring/ArchitectureAnalyzer.cs               │
  │        ├─ ScoreSpof / ScoreCoupling /          │
  │        │  ScoreScalabilityGap                   │
  │        │     → (score, RuleContribution[])      │
  │        ├─ findings + recommendations            │
  │        └─ criticality-weighted overall score    │
  └──────────────────────────────────────────────┘
            │
   ┌────────┴─────────┐
   ▼                  ▼
  API (Program.cs)   CLI (Program.cs)
  /v1/predict-        inventory.json
  architecture        [--json]

Projects

Project	Role
CloudSealed.ML.Engine	Contract models + the scoring core. No I/O, no web.
CloudSealed.ML.API	Minimal ASP.NET API: GET /health, POST /v1/predict-architecture, X-Api-Key auth, Swagger UI.
CloudSealed.ML.CLI	Runs the same analyzer on a local JSON inventory, human or --json output.
CloudSealed.ML.Tests	xUnit: rules in isolation, breakdown fidelity, weighted overall score, HTTP endpoint.

Why the boundaries are where they are

Weights are data, not code. Every risk weight and threshold is a named constant in RiskRules.cs with a one-line rationale, kept out of the branching logic in ArchitectureAnalyzer.cs. Retuning the model is editing a table, and the rationale travels with the number.

Scoring is traceable by construction. ScoreSpof, ScoreCoupling and ScoreScalabilityGap do not return a bare integer — they return the score and the list of RuleContributions that produced it. The response therefore carries a scoreBreakdown where sum(points) (capped at 100) reconstructs every riskScore exactly; a test enforces that identity so the explanation can never drift from the number. This is the property a black-box model cannot offer and is the reason the method is deterministic rules rather than a trained classifier (see METHODOLOGY.md).

Contract is additive-only. The response shape mirrors framework4d-predictive-ml-client.ts field-for-field. scoreBreakdown, engineVersion and method were added as new fields; older clients ignore them, so the contract is extended without a breaking version bump.

One core, two surfaces. API and CLI both call the same Analyze() — no second path that could diverge.