Skip to content

Use Cases

MPL addresses semantic governance challenges across regulated industries. Here are concrete scenarios where typed contracts, quality metrics, and audit trails provide value.

Financial Advisory

Scenario: An AI agent provides investment recommendations that must comply with fiduciary duty regulations.

Without MPL: The agent produces JSON responses with no enforceable schema. Compliance cannot verify that required disclaimers, risk ratings, and suitability assessments are present. Audit requires manual log review.

With MPL:

  • SType: org.finance.InvestmentRecommendation.v1 enforces required fields (risk_rating, suitability_score, disclaimers)
  • QoM Profile: qom-strict-argcheck with IC >= 0.97 ensures assertion compliance
  • Assertions: "risk_rating must be present", "disclaimer text must reference regulatory body"
  • Audit: Semantic hash + provenance provide tamper-evident trail for FCA/PRA compliance
result = await client.send(
    stype="org.finance.InvestmentRecommendation.v1",
    payload={
        "ticker": "AAPL",
        "action": "buy",
        "risk_rating": "moderate",
        "suitability_score": 0.85,
        "disclaimers": ["Past performance..."]
    }
)
assert result.qom_passed  # IC threshold met

Healthcare Patient Summaries

Scenario: An AI agent generates patient summaries from clinical notes. Outputs must comply with HIPAA and maintain clinical accuracy.

Without MPL: No enforcement that PHI is handled correctly. No measurement of whether clinical claims are grounded in source notes. No audit trail for data access patterns.

With MPL:

  • SType: org.health.PatientSummary.v1 with strict schema for clinical fields
  • QoM Metrics: Groundedness (G >= 0.95) ensures claims cite source notes
  • Policy Engine: Enforces PHI access rules and consent requirements
  • Provenance: Tracks which clinical notes were used as inputs

Enterprise Calendar Scheduling

Scenario: An AI scheduling agent creates and modifies calendar events across an organization.

Without MPL: Calendar payloads vary across integrations. An agent might create events with missing required fields (timezone, end time) that pass the API but produce incorrect behavior.

With MPL:

  • SType: org.calendar.Event.v1 enforces ISO 8601 timestamps, required fields, valid timezones
  • Schema Fidelity: Catches missing or malformed fields before the API call
  • Provenance: Tracks which user request led to each calendar modification
{
  "stype": "org.calendar.Event.v1",
  "payload": {
    "title": "Quarterly Review",
    "start": "2025-03-15T14:00:00Z",
    "end": "2025-03-15T15:00:00Z",
    "timezone": "America/New_York",
    "attendees": ["alice@corp.com", "bob@corp.com"]
  }
}

RAG Pipelines

Scenario: A Retrieval-Augmented Generation pipeline answers questions using a document corpus. Answers must cite sources and maintain factual accuracy.

Without MPL: No structured way to verify that retrieved documents are relevant or that generated answers are grounded in the retrieved content. Quality degrades silently.

With MPL:

  • SType: eval.rag.RAGQuery.v1 for queries, eval.rag.SearchResult.v1 for results
  • QoM Profile: Groundedness metric ensures answers cite sources
  • Determinism: DJ metric verifies answer stability under temperature variation
  • Assertions: Minimum relevance score thresholds for retrieved documents

Multi-Agent Task Planning

Scenario: A planner agent decomposes complex tasks and delegates subtasks to executor agents. Each step must be typed and tracked.

Without MPL: Task plans are ad-hoc JSON. Executor agents may receive malformed subtask specifications. No visibility into which steps succeeded or failed semantically.

With MPL:

  • STypes: org.agent.TaskPlan.v1, org.agent.ToolInvocation.v1, org.agent.ToolResult.v1
  • Handshake: Planner and executor negotiate compatible STypes and QoM levels
  • Provenance: Full transformation chain from original request to final result
  • QoM: Schema Fidelity ensures each subtask specification is well-formed
sequenceDiagram
    participant U as User
    participant P as Planner Agent
    participant E as Executor Agent
    participant T as Tool

    U->>P: "Book a flight and hotel for next week"
    P->>E: TaskPlan.v1 (steps: [search_flights, book_flight, search_hotels, book_hotel])
    E->>T: ToolInvocation.v1 (tool: search_flights, args: {...})
    T-->>E: ToolResult.v1 (results: [...])
    E-->>P: TaskPlan status update
    P-->>U: Completed plan with provenance

Supply Chain Document Processing

Scenario: AI agents process invoices, purchase orders, and shipping manifests across multiple vendors with different formats.

With MPL:

  • STypes: Standardized document types across vendors (org.supply.Invoice.v1, org.supply.PurchaseOrder.v1)
  • Schema Fidelity: Ensures all vendor documents conform to agreed schemas
  • Ontology Adherence: Validates business rules (quantities match, dates are sequential)
  • Audit Trails: Every document transformation is hash-chained for SOX compliance

Common Patterns

Across these use cases, MPL provides consistent value through:

Pattern Benefit
Schema enforcement at the protocol layer Catch errors before they reach downstream systems
Quality SLOs per interaction Measurable, enforceable quality guarantees
Typed error responses Faster debugging with semantic context
Provenance chains End-to-end auditability across agent hops
Policy enforcement Organizational rules enforced at runtime
Progressive adoption Start observing, then enforce when ready