US-07: Guardrail Strictness — Safety vs Latency vs User Experience

graph TD
    subgraph "Inference Team (Safety)"
        I["Block everything suspicious.<br/>One hallucinated price = lawsuit risk.<br/>One PII leak = regulatory action.<br/>False positives are safer<br/>than false negatives."]
    end

    subgraph "Performance Team"
        P["The guardrail pipeline adds<br/>100-500ms per request.<br/>That's 25% of our latency<br/>budget. Can we make some<br/>checks async?"]
    end

    subgraph "Cost Team"
        C["Guardrail compute costs $15K/month.<br/>Plus, every blocked response triggers<br/>a retry or fallback that<br/>costs MORE than the original.<br/>False positive rate of 5%<br/>means $22K in wasted retries."]
    end

    subgraph "Product Team"
        PT["Users hate 'I can't help with<br/>that' responses. Every false block<br/>is a lost engagement. Our CSAT<br/>drops 0.3 points for each<br/>unnecessary refusal."]
    end

    I ---|"Latency<br/>cost"| P
    P ---|"Compute<br/>cost"| C
    C ---|"User<br/>experience"| PT
    PT ---|"Safety<br/>risk"| I

    style I fill:#ff6b6b,stroke:#333,color:#000
    style P fill:#4ecdc4,stroke:#333,color:#000
    style C fill:#f9d71c,stroke:#333,color:#000
    style PT fill:#a29bfe,stroke:#333,color:#000

graph TD
    A["LLM Response"] --> B["Sync Guardrails<br/>(must pass before delivery)"]
    B --> C{"Pass?"}
    C -->|"Yes"| D["Deliver to User"]
    C -->|"No"| E["Block → Fallback Response"]

    D --> F["Async Guardrails<br/>(run after delivery)"]
    F --> G{"Pass?"}
    G -->|"Yes"| H["Log clean"]
    G -->|"No"| I["Flag for review<br/>Optional: retract + replace"]

    subgraph "Sync Checks — 100ms budget"
        S1["PII Detection<br/>20ms | Precision: 0.98"]
        S2["Price Validation<br/>30ms | Precision: 1.0"]
        S3["ASIN Existence<br/>30ms | Precision: 1.0"]
        S4["Toxicity Quick Scan<br/>20ms | Precision: 0.94"]
    end

    subgraph "Async Checks — No time budget"
        A1["Hallucination Scoring<br/>300ms | Precision: 0.87"]
        A2["Competitor Mention<br/>50ms | Precision: 0.96"]
        A3["Scope Drift<br/>200ms | Precision: 0.82"]
        A4["Full Quality Score<br/>400ms | Precision: 0.79"]
    end

    style B fill:#eb3b5a,stroke:#333,color:#fff
    style F fill:#fd9644,stroke:#333,color:#000
    style D fill:#2d8659,stroke:#333,color:#fff

graph LR
    subgraph "High Precision, Low Recall (Current)"
        HP["Blocks: 2% of responses<br/>False positives: 0.3%<br/>Missed violations: 1.5%<br/>✅ Users rarely frustrated<br/>⚠️ Some bad responses slip through"]
    end

    subgraph "High Recall, Low Precision (Aggressive)"
        HR["Blocks: 8% of responses<br/>False positives: 5%<br/>Missed violations: 0.2%<br/>❌ Users frequently frustrated<br/>✅ Almost nothing slips through"]
    end

    subgraph "Balanced (Target)"
        BAL["Blocks: 3.5% of responses<br/>False positives: 1%<br/>Missed violations: 0.8%<br/>✅ Acceptable frustration<br/>✅ Acceptable safety"]
    end

    style HP fill:#fd9644,stroke:#333,color:#000
    style HR fill:#eb3b5a,stroke:#333,color:#fff
    style BAL fill:#2d8659,stroke:#333,color:#fff

graph TD
    A["LLM says:<br/>'Demon Slayer Vol 1 is $9.99'"] --> B{"Catalog says<br/>price is $9.99?"}
    B -->|"Match"| C["✅ Pass"]
    B -->|"Mismatch"| D["Replace with catalog price"]
    B -->|"No price in catalog"| E["Remove price from response"]

    style C fill:#2d8659,stroke:#333,color:#fff
    style D fill:#fd9644,stroke:#333,color:#000
    style E fill:#eb3b5a,stroke:#333,color:#fff

graph TD
    subgraph "The Dilemma"
        D1["Real-time hallucination scoring<br/>takes 300ms and is only<br/>87% precise.<br/><br/>Async scoring catches issues<br/>after delivery but user already<br/>saw the hallucination."]
    end

    subgraph "Option A: Sync Scoring"
        A1["✅ Catches hallucinations before user sees them<br/>❌ Adds 300ms to every response<br/>❌ 13% false positive rate blocks good responses<br/>❌ $8K/month in compute"]
    end

    subgraph "Option B: Async Scoring"
        B1["✅ Zero latency impact<br/>✅ Zero false-positive UX impact<br/>❌ User sees hallucination for 2-5 seconds<br/>❌ Retraction UX is awkward"]
    end

    subgraph "Option C: Hybrid (Decision)"
        C1["Sync: lightweight NLI check (50ms)<br/>blocks obvious hallucinations<br/><br/>Async: full scoring (300ms)<br/>catches subtle hallucinations<br/>and retract if needed"]
    end

    style A1 fill:#eb3b5a,stroke:#333,color:#fff
    style B1 fill:#fd9644,stroke:#333,color:#000
    style C1 fill:#2d8659,stroke:#333,color:#fff

graph TD
    A["User: 'Tell me about<br/>the history of Japan'"] --> B{"Is this manga/Amazon<br/>related?"}
    B -->|"No"| C["Redirect:<br/>'I specialize in manga.<br/>Can I help you find<br/>a good manga about<br/>Japanese history?'"]
    B -->|"Borderline"| D["Answer briefly,<br/>then redirect"]
    B -->|"Yes"| E["Answer normally"]

    style C fill:#fd9644,stroke:#333,color:#000
    style D fill:#f9d71c,stroke:#333,color:#000
    style E fill:#2d8659,stroke:#333,color:#fff

graph TD
    A["False Positive<br/>(legitimate response blocked)"] --> B["Fallback Response Generated<br/>(+$0.008 for retry with template)"]
    A --> C["User Frustration<br/>(-0.3 CSAT points)"]
    A --> D["Potential Escalation<br/>($4-8 per human agent minute)"]
    A --> E["Engagement Loss<br/>(15% of blocked users leave session)"]

    style A fill:#eb3b5a,stroke:#333,color:#fff

graph TD
    A["Guardrail Dashboard"] --> B["Block Rate"]
    A --> C["False Positive Rate"]
    A --> D["False Negative Rate"]
    A --> E["Latency Impact"]

    B --> B1["Block rate by guardrail<br/>Block rate by intent<br/>Block rate trend (7d)"]
    C --> C1["FP rate by guardrail<br/>FP review queue<br/>FP examples for tuning"]
    D --> D1["Escaped violations caught by async<br/>User-reported issues<br/>Escalation root causes"]
    E --> E1["Sync check latency p50/p95<br/>Async check completion time<br/>Retry/fallback rate"]

    style A fill:#54a0ff,stroke:#333,color:#000