Governance Session Builder

This example instruments a three-step loan decisioning workflow using plain System.Diagnostics.Activity spans enriched with VeriProof governance attributes — no Semantic Kernel or AutoGen required.

Use this pattern when you control the LLM calls directly or when you need fine-grained per-step governance metadata.

Packages used

Veriproof.Sdk.Core — OTel exporter
Veriproof.Sdk.Annotations — typed governance Activity extensions

Prerequisites


dotnet add package Veriproof.Sdk.Core
dotnet add package Veriproof.Sdk.Annotations
dotnet add package Microsoft.Extensions.AI

Configure the exporter


// Program.cs
using OpenTelemetry.Trace;
using Veriproof.Sdk;
 
builder.Services.AddOpenTelemetry()
    .WithTraces(traces => traces
        .AddSource("LoanDecisioning")
        .AddVeriproofTracing(options =>
        {
            options.ApiKey        = Environment.GetEnvironmentVariable("VERIPROOF_API_KEY")!;
            options.ApplicationId = Environment.GetEnvironmentVariable("VERIPROOF_APP_ID")!;
        })
    );
 
builder.Services.AddSingleton<LoanDecisionService>();

Full example


// LoanDecisionService.cs
using System.Diagnostics;
using Microsoft.Extensions.AI;
using Veriproof.Sdk.Annotations;
using Veriproof.Sdk.Annotations.Models;
 
public sealed class LoanDecisionService(IChatClient chatClient)
{
    private static readonly ActivitySource Tracer = new("LoanDecisioning");
 
    public async Task<LoanDecisionResult> DecideAsync(
        LoanApplication  application,
        CancellationToken ct = default)
    {
        // ── Session span ────────────────────────────────────────────────────────
        using var session = Tracer.StartActivity("ai.session");
        session?
            .SetSessionIntent(SessionIntent.transaction)
            .SetTransactionId($"loan:{application.Id}")
            .SetDataSensitivity(DataSensitivity.Restricted)
            .SetRegulatoryScope("ECOA")
            .SetHumanReviewRequired(false);
 
        // ── Step 1: Classify ─────────────────────────────────────────────────
        string riskTier;
        using (var step1 = Tracer.StartActivity("ai.step"))
        {
            step1?
                .SetStepType(StepType.llm_call)
                .SetStepName("classify_risk_tier")
                .SetGroundingSource("credit-bureau-v2");
 
            var classifyPrompt = new[]
            {
                new ChatMessage(ChatRole.System,
                    "Classify the loan applicant into LOW_RISK, MEDIUM_RISK, or HIGH_RISK " +
                    "based on credit score and debt-to-income ratio. Reply with only the tier."),
                new ChatMessage(ChatRole.User,
                    $"Credit score: {application.CreditScore}. " +
                    $"DTI ratio: {application.DebtToIncomeRatio:P0}.")
            };
 
            var classifyResponse =
                await chatClient.GetResponseAsync(classifyPrompt, ct: ct);
 
            riskTier = classifyResponse.Message.Text?.Trim() ?? "HIGH_RISK";
            step1?.SetStepOutcome(StepOutcome.success);
            step1?.SetGroundingStatus(GroundingStatus.grounded);
        }
 
        // ── Step 2: Assess eligibility ────────────────────────────────────────
        string eligibilitySummary;
        using (var step2 = Tracer.StartActivity("ai.step"))
        {
            step2?
                .SetStepType(StepType.llm_call)
                .SetStepName("assess_eligibility")
                .SetStepSequence(2);
 
            var assessPrompt = new[]
            {
                new ChatMessage(ChatRole.System,
                    "You are a loan eligibility officer. " +
                    "Given the risk tier, recommend APPROVE, CONDITIONAL, or DENY " +
                    "and provide a brief one-sentence justification."),
                new ChatMessage(ChatRole.User,
                    $"Risk tier: {riskTier}. " +
                    $"Requested amount: {application.RequestedAmount:C0}.")
            };
 
            var assessResponse =
                await chatClient.GetResponseAsync(assessPrompt, ct: ct);
 
            eligibilitySummary = assessResponse.Message.Text ?? "DENY";
            step2?.SetStepOutcome(StepOutcome.success);
        }
 
        // ── Step 3: Final decision ────────────────────────────────────────────
        bool approved;
        using (var step3 = Tracer.StartActivity("ai.step"))
        {
            step3?
                .SetStepType(StepType.tool_call)
                .SetStepName("record_decision")
                .SetStepSequence(3);
 
            approved = eligibilitySummary.Contains("APPROVE",
                StringComparison.OrdinalIgnoreCase);
 
            step3?.SetStepOutcome(approved ? StepOutcome.success : StepOutcome.failure);
        }
 
        // ── Enrich session with outcome ───────────────────────────────────────
        var riskLevel = riskTier == "HIGH_RISK"   ? RiskLevel.High
                      : riskTier == "MEDIUM_RISK" ? RiskLevel.Medium
                      :                              RiskLevel.Low;
 
        session?
            .SetRiskLevel(riskLevel)
            .SetSessionOutcome(approved ? SessionOutcome.success : SessionOutcome.failure)
            .SetGovernanceDecision(
                DecisionContext.WithOptions(
                    prompt:       $"Loan application {application.Id}",
                    decision:     approved ? "APPROVE" : "DENY",
                    decisionType: DecisionType.Approval,
                    options:      new[] { "APPROVE", "CONDITIONAL", "DENY" }
                )
            );
 
        if (riskLevel >= RiskLevel.Medium)
        {
            session?.AddRiskFactor(
                RiskFactor.LowConfidence(
                    $"Risk tier {riskTier} — additional review recommended"));
        }
 
        return new LoanDecisionResult(
            ApplicationId: application.Id,
            Approved:      approved,
            RiskTier:      riskTier,
            Summary:       eligibilitySummary
        );
    }
}
 
public record LoanApplication(
    string  Id,
    int     CreditScore,
    double  DebtToIncomeRatio,
    decimal RequestedAmount
);
 
public record LoanDecisionResult(
    string Id,
    bool   Approved,
    string RiskTier,
    string Summary
) {
    public LoanDecisionResult(string ApplicationId, bool Approved, string RiskTier, string Summary)
        : this(ApplicationId, Approved, RiskTier, Summary) {}
}

Span hierarchy


ai.session  (LoanDecisioning — loan:{Id})
│  veriproof.session.intent = "transaction"
│  veriproof.risk.level     = "medium"
│  veriproof.governance.decision.type = "approval"
│
├── ai.step  classify_risk_tier
│     veriproof.step.type    = "llm_call"
│     veriproof.step.outcome = "success"
│
├── ai.step  assess_eligibility
│     veriproof.step.type    = "llm_call"
│     veriproof.step.outcome = "success"
│
└── ai.step  record_decision
      veriproof.step.type    = "tool_call"
      veriproof.step.outcome = "success"

What you’ll see in VeriProof

Field	Value
Session intent	`transaction`
Risk level	Derived from credit score
Governance decision	`APPROVE` / `DENY` with prompt and options
Step types	`llm_call`, `llm_call`, `tool_call`
Data sensitivity	`Restricted`
Regulatory scope	`ECOA`

Next steps

AutoGen multi-agent example — orchestrate multiple agents with AutoGenOTelAdapter
Semantic Kernel example — automatic span capture with VeriproofKernelFilter
Veriproof.Sdk.Annotations reference — full enum and extension method reference