Architecture Atlas

This atlas turns the book’s arguments into diagrams. Use it as a visual checklist when designing or reviewing a production AI system.

The Production Boundary

flowchart LR
    User["User or operator"] --> Intake["Input boundary"]
    Intake --> Evidence["Evidence layer"]
    Evidence --> Prompt["Prompt and context builder"]
    Prompt --> Model["Model call"]
    Model --> Validate["Validation and eval checks"]
    Validate --> Human["Human review"]
    Human --> State["Typed workflow state"]
    State --> Audit["Audit packet"]
    State --> Observe["Observability"]
    Observe --> Improve["Eval and improvement loop"]
    Improve --> Prompt

The model is one component inside a controlled workflow. The system owns evidence, state, review, audit, and improvement.

Evaluation Loop

flowchart TD
    Change["Prompt, model, retrieval, or code change"] --> Fixtures["Golden and adversarial fixtures"]
    Fixtures --> Run["Eval run"]
    Run --> Score["Risk-weighted score"]
    Score --> Gate{"Release gate"}
    Gate -->|pass| Deploy["Deploy"]
    Gate -->|fail| Fix["Fix source of regression"]
    Deploy --> Monitor["Production monitoring"]
    Monitor --> Corrections["Human corrections and incidents"]
    Corrections --> Fixtures

Evaluation is not a one-time benchmark. Production corrections should feed the fixture set.

Typed Workflow

stateDiagram-v2
    [*] --> Open
    Open --> WaitingForDocuments: DocumentUploaded
    WaitingForDocuments --> WaitingForDocuments: MissingDataDetected
    WaitingForDocuments --> ReadyForAnalystReview: ExtractionSucceeded
    ReadyForAnalystReview --> ApprovedByHuman: AnalystApproved
    ReadyForAnalystReview --> RejectedByHuman: AnalystRejected
    ApprovedByHuman --> [*]
    RejectedByHuman --> [*]

The important absence is as meaningful as the arrows: there is no ModelApproved transition.

Human-Control Ladder

flowchart BT
    Decide["AI decides"] --> Validate["AI validates"]
    Validate --> Route["AI routes"]
    Route --> Draft["AI drafts"]
    Draft --> Recommend["AI recommends"]
    Recommend --> Prepare["AI prepares"]

Risk should push the system downward toward preparation and recommendation, with human-owned transitions for consequential decisions.

Observability Records

flowchart LR
    Workflow["Workflow execution"] --> Debug["Debug logs"]
    Workflow --> Trace["Traces and spans"]
    Workflow --> Semantic["Semantic AI events"]
    Workflow --> Audit["Audit records"]

    Debug --> Engineer["Engineer debugging"]
    Trace --> SRE["Operations and latency"]
    Semantic --> Eval["Evaluation and drift"]
    Audit --> Compliance["Compliance and accountability"]

Do not force one record type to serve every audience. Debugging, operations, evaluation, and audit have different needs.

Security Boundary

flowchart TD
    Trusted["Trusted policy and developer instructions"] --> Builder["Prompt/context builder"]
    User["Authenticated user request"] --> Builder
    Docs["Retrieved documents as untrusted evidence"] --> Label["Evidence labeling"]
    Tools["Tool results as untrusted evidence"] --> Label
    Label --> Builder
    Builder --> Model["Model"]
    Model --> Proposal["Proposal or draft"]
    Proposal --> Policy["Policy and permission checks"]
    Policy -->|low risk| Action["Allowed action"]
    Policy -->|high risk| Review["Human approval"]
    Policy -->|forbidden| Block["Blocked"]

The core rule is authority separation: evidence is not instruction, and proposal is not decision.

Economics Routing

flowchart TD
    Task["Workflow task"] --> Rules{"Can deterministic code solve it?"}
    Rules -->|yes| Code["Use code or database constraints"]
    Rules -->|no| Risk{"High risk or high ambiguity?"}
    Risk -->|low| Small["Small or medium model"]
    Risk -->|high| Frontier["Frontier model plus review"]
    Small --> Cache{"Safe to cache?"}
    Frontier --> Review["Human review"]
    Cache -->|yes| Cached["Cache with version and tenant rules"]
    Cache -->|no| Direct["Run uncached"]

Model routing is not only cost optimization. It is risk routing.

Tool Permission Matrix

flowchart LR
    Model["Model"] --> Read["Read scoped evidence"]
    Model --> Draft["Draft internal note"]
    Model -. blocked .-> External["External communication"]
    Model -. blocked .-> Final["Final case decision"]
    External --> Human["Human approval"]
    Final --> Human
    Human --> Audit["Audit log"]

The model may prepare and draft. High-risk writes route through human approval and audit.

Capstone Flow

flowchart TD
    Open["Case opened"] --> Upload["Document uploaded"]
    Upload --> Outbox["Outbox extraction request"]
    Outbox --> Extract["Extraction worker"]
    Extract --> Evidence["Versioned evidence packet"]
    Evidence --> AI["AI draft and risk signal"]
    AI --> InlineEval["Inline eval and policy checks"]
    InlineEval --> Queue["Analyst review queue"]
    Queue --> Decision{"Human decision"}
    Decision -->|approve| Approved["ApprovedByHuman"]
    Decision -->|reject| Rejected["RejectedByHuman"]
    Decision -->|more evidence| Upload
    Approved --> Audit["Audit packet finalized"]
    Rejected --> Audit
    Audit --> Metrics["Metrics, traces, eval candidates"]

The capstone combines every pillar: evaluation, typed state, human control, observability, security, economics, and distribution-grade trust artifacts.