# Verdict - Courtroom Simulation A WPF desktop application for modeling jury verdict outcomes using AI agents. The application simulates a courtroom environment where AI-powered agents (jurors, judge, lawyers, witnesses) interact and form opinions based on trial events, evidence, and testimony. ## Overview Verdict is a .NET 9 WPF application designed for native Windows deployment. It uses an MVVM architecture and supports multiple LLM providers (OpenAI, Anthropic, Google Gemini, etc.) to power AI agent behavior in a courtroom simulation. ### Key Features - **Dynamic Courtroom Layout**: Full courtroom with Judge, Witness Stand, Court Reporter, Jury Box (12 seats in 3x4 formation), Counsel Tables, and Gallery - **AI Agent System**: Configurable AI agents with demographic backgrounds, biases, sentiment tracking, and verdict leanings - **Transcript Processing**: Load and process court transcripts that influence agent opinions - **Evidence Management**: Admit evidence documents that create trial memories and shift agent opinions - **Photo Evidence**: Admit and describe photo evidence through the clerk, with image file support (.jpg, .png, .gif, .bmp) - **Exhibit List**: List view of all evidence with exhibit numbers, summaries, and clerk descriptions - **Agent Chat**: Right-click chat with any agent using their configured LLM, with conversation memory and fallback role-based responses - **Memory Wipe**: Wipe chat memories per-agent from the chat window - **Counsel Discussion**: Simulate a sidebar discussion between defense and prosecution lawyers, with judge ruling - **Client Consultation**: Privileged attorney-client conversation window with settlement authority tracking - **Juror Report**: Right-click sentiment and memory report for jurors - **Model Weights**: Configure bias factor weights (age, gender, education, income, political, ethnicity, religion, experience, legal knowledge, professional background, community ties) via the Model Weights window - **Trial Stage Menu**: Top-level Stage menu for trial phase (Discovery, Pretrial, Trial) and court phase selection - **Dynamic Title Bar**: Title shows case name, parties, and trial stage (e.g., "VERDICT - Plaintiff v. Defendant - (TRIAL)") - **Multi-Provider Support**: Connect to various LLM providers (OpenAI, Anthropic, Google Gemini, Ollama, DeepSeek, Grok, etc.) - **Case File System**: Save/load simulations (.jur files) with full state preservation - **Character Profiles**: Save reusable agent configurations as .vcs files - **PDF Reports**: Generate PDF case reports via QuestPDF - **Dynamic Role Assignment**: Automatically assign agents to recommended roles based on case analysis - **Logger**: File-based logging to `VerdictSimulation.log` in the application directory ## Architecture ``` Verdict/ ├── Models/ # Data entities (Agent, CaseFile, EvidenceDocument, etc.) ├── ViewModels/ # MVVM ViewModels (MainViewModel, CaseSettingsViewModel, ModelSettingsViewModel) ├── Views/ # XAML Windows and UserControls ├── Services/ # Business logic — 17 dedicated services ├── Providers/ # LLM provider implementations ├── Resources/ # LegalDatabase.json ├── Help/ # HTML help pages └── Tests/ # Console test harness ``` ### Core Models | Model | Purpose | |-------|---------| | `Agent` | Represents an individual in the courtroom (juror, judge, lawyer, etc.) with personality, memories, and verdict lean | | `CaseFile` | Complete simulation state including agents, evidence, event log, and legal framework | | `MemoryEntry` | A single memory or trial observation with timestamp and strength | | `CourtroomEvent` | An event that occurred in the courtroom, visible to specific roles | | `EvidenceDocument` | Evidence admitted into the case with summary and detailed LLM analysis | | `AIModelConfiguration` | LLM provider configuration (API key, model ID, endpoint) | | `JuryInstructions` | Legal instructions given to the jury | | `VerdictForm` | The verdict form for recording jury decisions | | `ExtractedEntities` | Entities extracted from transcripts (charges, witnesses, evidence) | | `AgentInteraction` | Agent messages, legal citations, and interaction results | | `Charge` / `CauseOfAction` / `Element` | Legal charge modeling | | `BiasFactor` | Demographic bias factor with configurable weight | ### Agent Roles ```csharp public enum AgentRole { Judge, // Presides over the trial Witness, // Provides testimony Reporter, // Records proceedings Juror, // Decides the verdict (12 seats) AlternateJuror, // Replacement juror (2 seats) Lawyer, // Represents prosecution or defense Client, // The plaintiff or defendant Observer // Gallery observer } ``` ### Service Architecture The application follows a **delegated service pattern** — `MainViewModel` owns the agent collections and case state, but delegates all domain logic to dedicated injected services. | Service | Interface | Responsibility | |---------|-----------|----------------| | `CaseService` | `ICaseService` | Save/load .jur case files; generate PDF reports | | `TranscriptService` | `ITranscriptService` | Load/parse transcript files; LLM entity extraction | | `SettingsService` | `ISettingsService` | Default case settings persistence | | `CourtroomManagerService` | `ICourtroomManagerService` | Courtroom initialization, slot occupancy, agent reset | | `EvidenceAnalysisService` | `IEvidenceAnalysisService` | Evidence strength assessment, damage estimation, exposure calculation, LLM document analysis | | `DebateService` | `IDebateService` | Stage advancement, influence calculation, event broadcasting | | `JuryCalculationService` | `IJuryCalculationService` | Average lean, likely verdict string, opinion influence logic | | `CaseEntityMapper` | `ICaseEntityMapper` | Maps extracted entities (charges/evidence/witnesses) onto CaseFile | | `AgentAssignmentService` | `IAgentAssignmentService` | Case analysis, dynamic role assignment, default agent generation | | `JuryDemographicsService` | `IJuryDemographicsService` | Generates jury panels based on county/state demographics | | `AgentInteractionService` | `IAgentInteractionService` | Manages agent statements, examination, deliberation, court record internalization | | `CharacterManager` | — | Save/load .vcs character profiles | | `LegalDatabaseService` | — | Query legal citations (IPC sections, U.S. Federal Law) | | `ReportGenerationService` | `IReportGenerationService` | PDF report rendering with QuestPDF | | `ProviderDiscoveryService` | — | Discovers installed LLM providers and their metadata | | `ModelDownloadService` | — | Downloads ONNX models from HuggingFace for local inference | | `Logger` | — | Static file-based logger writing to `VerdictSimulation.log` | ### Views | Window | Purpose | |--------|---------| | `MainWindow` | Primary courtroom UI | | `AgentProfileWindow` | Edit agent demographics, personality, and model settings | | `AgentChatWindow` | LLM-powered chat with any agent; memory persistence and wipe | | `CaseSettingsWindow` | Edit case name, parties, jurisdiction, and legal framework | | `ModelSettingsWindow` | Manage LLM provider configurations | | `ModelWeightsWindow` | Configure bias factor weights for jury modeling | | `JurorReportWindow` | Sentiment, verdict lean, and memory report for a juror | | `JurorConversationWindow` | Targeted questioning of a juror | | `CounselDiscussionWindow` | Sidebar discussion between defense and prosecution with judge ruling | | `ClientConversationWindow` | Privileged attorney-client consultation | | `ExhibitListWindow` | Grid view of all admitted evidence | | `AddModelDialog` | Add a new LLM model configuration | | `ModelSelectionDialog` | Select a model from available configurations | | `ModelDetailsDialog` | View detailed model configuration | ### LLM Providers | Provider | API Key Link | Models Supported | |----------|-------------|-------------------| | **OpenAI** | [platform.openai.com/api-keys](https://platform.openai.com/api-keys) | GPT-4o, GPT-4 Turbo, GPT-4 | | **Anthropic** | [console.anthropic.com/api-keys](https://console.anthropic.com/api-keys) | Claude 3.5 Sonnet, Claude 3 Opus | | **Google Gemini** | [aistudio.google.com/app/apikey](https://aistudio.google.com/app/apikey) | Gemini 1.5 Pro, Gemini 1.5 Flash | | **Ollama** | [ollama.com](https://ollama.com/download) | Llama 3, Mistral, Phi3 (local) | | **Hugging Face (GGUF)** | [huggingface.co/settings/tokens](https://huggingface.co/settings/tokens) | GGUF models via LLamaSharp | | **ONNX** | [huggingface.co/models](https://huggingface.co/models?library=onnx) | ONNX models via Microsoft.ML.OnnxRuntime | | **Alibaba (Qwen)** | [dashscope.console.aliyun.com](https://dashscope.console.aliyun.com) | Qwen 2, Qwen 1.5 | | **NVIDIA NIM** | [build.nvidia.com](https://build.nvidia.com) | Mixtral, Llama 3 (via NIM) | | **Intel Gaudi** | [vault.habana.ai](https://vault.habana.ai) | Gaudi-accelerated models | | **DeepSeek** | [platform.deepseek.com](https://platform.deepseek.com) | DeepSeek-V3, DeepSeek-R1 | | **Grok (xAI)** | [console.x.ai](https://console.x.ai) | Grok Beta | #### Quick Start - Adding a Model 1. Click **"Add Model"** button in the Models panel 2. Select your provider from the dropdown 3. Enter your API key from the link above 4. Choose a model ID (e.g., `gpt-4o`, `claude-3-5-sonnet-20240620`) 5. Click **"Test Connection"** to verify 6. Click **"Save"** to add to your available models ## Test Harness The project includes a comprehensive test harness in the `Tests/` directory — a console application that performs unit and integration tests on all core components. ### Architecture | File | Type | Purpose | |------|------|---------| | `Program.cs` | Entry Point | Launches the test harness | | `BaseTestClass.cs` | Base Class | Shared `Assert()` helper + pass/fail counters | | `TestHarness.cs` | Legacy Runner | Original monolithic test runner (still functional) | | `TestRunner.cs` | New Runner | Aggregates results from ModelTests + ServiceTests + ProviderTests + ViewModelTests | | `ModelTests.cs` | Test Suite | 12 tests covering all 20 model classes | | `ServiceTests.cs` | Test Suite | 15 tests covering all services (SettingsService has full round-trip tests) | | `ProviderTests.cs` | Test Suite | 8 tests covering provider discovery + OnnxProvider + HuggingFaceProvider | | `ViewModelTests.cs` | Test Suite | 3 placeholder tests for MainViewModel, CaseSettingsViewModel, ModelSettingsViewModel | ### Running Tests ```bash dotnet run --project Tests/Tests.csproj ``` **Note**: LLM response generation tests are skipped by default as they require live API keys. ## Agent System ### Verdict Lean System Each agent with voting capability (`HasOpinion`) tracks a `VerdictLean` value: - `0.0` = Strongly favors Defendant - `0.5` = Undecided - `1.0` = Strongly favors Prosecution/Plaintiff The lean updates based on transcript influence, admitted evidence, and peer influence during deliberation. ### Memory System Agents maintain two types of memories: - **Memories**: Background life experiences, identity traits, and chat/consultation exchanges (tagged by `Source`: `"Chat"`, `"CounselDiscussion"`, `"ClientConsultation"`) - **TrialEvents**: Specific observations during the current trial Both use `MemoryEntry` with `Content`, `Timestamp`, `Strength`, `Source`, and `IsFromDocument`. ### Sentiment Analysis Heuristic sentiment analysis on trial content: - Objection/Overruled → Negative shift - Sustained → Positive shift - Lied/False/Inconsistent → Significant negative shift - Expert/DNA/Fact → Positive shift ### Bias System Each agent has a `Bias` value (-1.0 to 1.0) that amplifies emotional responses to bias-confirming information and modifies verdict lean updates. Bias factor weights are configurable per demographic dimension via the Model Weights window. ## Data Formats ### Case Files (.jur) JSON files storing complete simulation state: ```json { "CaseName": "Sample Case", "CaseNumber": "23-CV-1234", "CourtName": "Superior Court", "Mode": "Civil", "Jurisdiction": "State", "JurorCount": 12, "Agents": [...], "Evidence": [...], "EventLog": [...], "Instructions": {...}, "Verdict": {...} } ``` ### Character Profiles (.vcs) Verdict Character Settings — reusable agent configurations stored in `/Characters` directory. ### Assets Each simulation creates a sibling directory (`name_Assets/`) containing evidence documents and agent artifacts. ## Build and Development ### Prerequisites - .NET 9 SDK - Windows 11 (for WPF) ### Commands ```bash dotnet build dotnet run dotnet run --project Tests/Tests.csproj dotnet clean ``` ## Courtroom Layout ``` +--------------------------------------------------+ | GALLERY (4 obs + 2 clients) | | +----+----+----+----+----+----+----+----+----+ | | |Alt1|Alt2|Obs1|Obs2|Obs3|Obs4|Plnt|DefC|Clie| | | +----+----+----+----+----+----+----+----+----+ | | | | +----------+ +----------+| | | JURY BOX | [WITNESS] [COURT | || | | 3x4 | STAND REPORTER] | OBSERVER || | | Jurors | | 6 || | +----------+ +----------+| | | | [DEFENSE TABLE] [PODIUM] [PROSECUTION] | | Lawyer + Client Lawyer + Client | | | | [ JUDGE ] | +--------------------------------------------------+ ``` ### Seating Capacity - **Judge Area**: 1 Judge, 1 Witness, 1 Reporter - **Jury Box**: 12 Jurors (3x4 formation) + 2 Alternate Jurors - **Gallery**: 4 Observers + 2 Client slots - **Counsel Tables**: Defense (Left) and Prosecution/Plaintiff (Right) ## Court Phases ```csharp public enum CourtPhase { CaseGeneration, CourtOpening, LegalResearch, OpeningStatements, WitnessTestimony, CrossExamination, PartyStatements, ClosingArguments, JuryInstructions, JuryDeliberation, VerdictAnnouncement, CourtClosing } ``` Use the **Stage** menu to switch between trial phases (Discovery, Pretrial, Trial) and court phases. The current stage is shown in the status bar with a **"→ Next"** button to advance. ## License This project is for demonstration and research purposes. # Verdict